test-nmmin.cpp
#include <RcppArmadillo.h>
#include <vector>
#include <map>
#include "c_optim.h"
namespace rstpm2 {
using namespace Rcpp;
using namespace arma;
// typedefs
typedef bool constraintfn(int, double *, void *);
// Hadamard element-wise multiplication for the _columns_ of a matrix with a vector
mat rmult(mat m, vec v) {
mat out(m);
out.each_col() %= v;
return out;
}
mat lmult(vec v, mat m) {
return rmult(m,v);
}
// element-wise multiplication and division for NumericVector and arma::vec
NumericVector ew_mult(NumericVector nv, vec v) {
return as<NumericVector>(wrap(as<vec>(wrap(nv)) % v));
}
NumericVector ew_div(NumericVector nv, vec v) {
return as<NumericVector>(wrap(as<vec>(wrap(nv)) / v));
}
void Rprint(NumericMatrix m) {
for (int i=0; i<m.nrow(); ++i) {
for (int j=0; j<m.ncol(); ++j)
Rprintf("%f ", m(i,j));
Rprintf("\n");
}
}
void Rprint(NumericVector v) {
for (int i=0; i<v.size(); ++i)
Rprintf("%f ", v(i));
Rprintf("\n");
}
void Rprint(vec v) {
for (size_t i=0; i<v.size(); ++i)
Rprintf("%f ", v(i));
Rprintf("\n");
}
void Rprint(rowvec v) {
for (size_t i=0; i<v.size(); ++i)
Rprintf("%f ", v(i));
Rprintf("\n");
}
void Rprint(mat m) {
for (size_t i=0; i<m.n_rows; ++i) {
for (size_t j=0; j<m.n_cols; ++j)
Rprintf("%f ", m(i,j));
Rprintf("\n");
}
}
vec pnorm01(vec x) {
vec out(x.size());
for (size_t i=0; i<x.size(); ++i)
out(i) = R::pnorm(x(i),0.0,1.0,1,0);
return out;
//return as<vec>(wrap(pnorm(as<NumericVector>(wrap<vec>(x)),0.0,1.0)));
}
vec qnorm01(vec x) {
vec out(x.size());
for (size_t i=0; i<x.size(); ++i)
out(i) = R::qnorm(x(i),0.0,1.0,1,0);
return out;
}
vec dnorm01(vec x) {
vec out(x.size());
for (size_t i=0; i<x.size(); ++i)
out(i) = R::dnorm(x(i),0.0,1.0,0);
return out;
}
vec logit(vec p) {
return log(p/(1-p));
}
vec expit(vec x) {
return 1/(1+exp(-x));
}
RcppExport SEXP tapplySum(SEXP s_y, SEXP s_group) {
NumericVector y(s_y);
NumericVector group(s_group);
NumericVector::iterator iy, igroup;
std::map<double,double> out;
for (iy = y.begin(), igroup = group.begin(); iy != y.end(); ++iy, ++igroup) {
out[*igroup] += *iy;
}
return wrap(out);
}
// Q: do we need parscale, trace, reltol and criterion in these structs?
class stpm2 {
public:
stpm2(SEXP sexp) {
List list = as<List>(sexp);
init = as<NumericVector>(list["init"]);
X = as<mat>(list["X"]);
XD = as<mat>(list["XD"]);
bhazard = as<vec>(list["bhazard"]);
wt = as<vec>(list["wt"]);
event = as<vec>(list["event"]);
time = as<vec>(list["time"]);
delayed = as<int>(list["delayed"]);
X0.set_size(1,1); X0.fill(0.0);
wt0.set_size(1); wt0.fill(0.0);
if (delayed == 1) {
X0 = as<mat>(list["X0"]); // optional
XD0 = as<mat>(list["XD0"]); // optional
wt0 = as<vec>(list["wt0"]); // optional
}
parscale = as<vec>(list["parscale"]);
reltol = as<double>(list["reltol"]);
kappa = as<double>(list["kappa"]);
trace = as<int>(list["trace"]);
}
// defaults are for the PH model
virtual vec link(vec S) { return log(-log(S)); }
virtual vec ilink(vec x) { return exp(-exp(x)); }
virtual vec h(vec eta, vec etaD) { return etaD % exp(eta); }
virtual vec H(vec eta, vec etaD) { return exp(eta); }
virtual mat gradH(vec eta, vec etaD, mat X, mat XD) { return rmult(X,exp(eta)); }
virtual mat gradh(vec eta, vec etaD, mat X, mat XD) {
return rmult(XD, exp(eta)) + rmult(X, etaD % exp(eta));
}
virtual double negll(NumericVector beta);
virtual void grad_negll(NumericVector beta);
NumericVector init;
mat X, XD, X0, XD0;
vec bhazard,wt,wt0,event,time,parscale;
double reltol, kappa;
int delayed, trace;
};
double stpm2::negll(NumericVector beta) {
int n = init.size();
vec vbeta(&beta[0],n);
vbeta = vbeta % parscale;
vec eta = X * vbeta;
vec etaD = XD * vbeta;
vec h = this->h(eta,etaD) + bhazard;
vec H = this->H(eta,etaD) + bhazard;
double constraint = kappa/2.0 * sum(h % h % (h<0)); // sum(h^2 | h<0)
vec eps = h*0.0 + 1.0e-16;
h = max(h,eps);
double ll = sum(wt % event % log(h)) - sum(wt % H) - constraint;
if (delayed == 1) {
vec eta0 = X0 * vbeta;
vec etaD0 = XD0 * vbeta;
mat H0 = this->H(eta0,etaD0);
ll += sum(wt0 % H0);
}
return -ll;
}
void stpm2::grad_negll(NumericVector beta) {
int n = beta.size();
vec vbeta = as<vec>(wrap(beta));
vbeta = vbeta % parscale;
vec eta = X * vbeta;
vec etaD = XD * vbeta;
vec h = this->h(eta,etaD) + bhazard;
// vec H = exp(eta);
mat gradH = this->gradH(eta,etaD,X,XD);
mat gradh = this->gradh(eta,etaD,X,XD);
mat Xgrad = -gradH + rmult(gradh, event / h);
mat Xconstraint = - kappa*rmult(gradh,h);
vec eps = h*0.0 + 1.0e-16; // hack
//Xgrad.rows(h<=eps) = Xconstraint.rows(h<=eps);
Xgrad = rmult(Xgrad,wt);
rowvec vgr = sum(Xgrad,0);
if (delayed == 1) {
vec eta0 = X0 * vbeta;
vec etaD0 = XD0 * vbeta;
mat gradH0 = this->gradH(eta0, etaD0, X0, XD0);
vgr += sum(rmult(gradH0, wt0),0);
}
for (int i = 0; i<n; ++i) {
// gr[i] = -vgr[i]*parscale[i];
}
}
class stpm2PO : public stpm2 {
public:
stpm2PO(SEXP sexp) : stpm2(sexp) { }
vec link(vec S) { return -logit(S); }
vec ilink(vec x) { return expit(-x); }
vec h(vec eta, vec etaD) { return etaD % exp(eta) % expit(-eta); }
vec H(vec eta, vec etaD) { return log(1+exp(eta)); }
mat gradH(vec eta, vec etaD, mat X, mat XD) {
return rmult(X,exp(eta) % expit(-eta));
}
mat gradh(vec eta, vec etaD, mat X, mat XD) {
return rmult(X, etaD % exp(eta) % expit(-eta)) -
rmult(X, exp(2*eta) % etaD % expit(-eta) % expit(-eta)) +
rmult(XD, exp(eta) % expit(-eta));
}
};
class stpm2Probit : public stpm2 {
public:
stpm2Probit(SEXP sexp) : stpm2(sexp) { }
vec link(vec S) { return -qnorm01(S); }
vec ilink(vec eta) { return pnorm01(-eta); }
vec H(vec eta, vec etaD) { return -log(pnorm01(-eta)); }
vec h(vec eta, vec etaD) { return etaD % dnorm01(eta) / pnorm01(-eta); }
mat gradH(vec eta, vec etaD, mat X, mat XD) {
return rmult(X, dnorm01(eta) / pnorm01(-eta));
}
mat gradh(vec eta, vec etaD, mat X, mat XD) {
return rmult(X, -eta % dnorm01(eta) % etaD / pnorm01(-eta)) +
rmult(X, dnorm01(eta) % dnorm01(eta) / pnorm01(-eta) / pnorm01(-eta) % etaD) +
rmult(XD,dnorm01(eta) / pnorm01(-eta));
}
};
class stpm2AH : public stpm2 {
public:
stpm2AH(SEXP sexp) : stpm2(sexp) { }
vec link(vec S) { return -log(S); }
vec ilink(vec x) { return exp(-x); }
vec h(vec eta, vec etaD) { return etaD; }
vec H(vec eta, vec etaD) { return eta; }
mat gradh(vec eta, vec etaD, mat X, mat XD) { return XD; }
mat gradH(vec eta, vec etaD, mat X, mat XD) { return X; }
};
struct SmoothLogH {
int first_para, last_para;
mat S;
};
struct SmoothHaz {
mat X0, X1, X2, X3;
vec w;
double lambda;
};
template<class Smooth>
std::vector<Smooth> read_smoothers(List lsmooth) {}
template<>
std::vector<SmoothLogH> read_smoothers(List lsmooth) {
std::vector<SmoothLogH> smooth;
for(int i=0; i<lsmooth.size(); ++i) {
List lsmoothi = as<List>(lsmooth[i]);
List lsmoothiS = as<List>(lsmoothi["S"]);
SmoothLogH smoothi = {
as<int>(lsmoothi["first.para"]) - 1,
as<int>(lsmoothi["last.para"]) - 1,
as<mat>(lsmoothiS[0])
};
smooth.push_back(smoothi);
}
return smooth;
}
template<>
std::vector<SmoothHaz> read_smoothers(List lsmooth) {
std::vector<SmoothHaz> smooth;
for(int i=0; i<lsmooth.size(); ++i) {
List lsmoothi = as<List>(lsmooth[i]);
SmoothHaz smoothi = {
as<mat>(lsmoothi["X0"]),
as<mat>(lsmoothi["X1"]),
as<mat>(lsmoothi["X2"]),
as<mat>(lsmoothi["X3"]),
as<vec>(lsmoothi["w"]),
as<double>(lsmoothi["lambda"])
};
smooth.push_back(smoothi);
}
return smooth;
}
template<class Smooth, class Stpm2 = stpm2>
class pstpm2 : public Stpm2 {
public:
pstpm2(SEXP sexp) : Stpm2(sexp) {
List list = as<List>(sexp);
List lsmooth = as<List>(list["smooth"]);
sp = as<NumericVector>(list["sp"]);
reltol_search = as<double>(list["reltol_search"]);
// reltol = as<double>(list["reltol"]);
alpha = as<double>(list["alpha"]);
criterion = as<int>(list["criterion"]);
smooth = read_smoothers<Smooth>(lsmooth);
}
NumericVector sp;
double alpha, reltol_search;
int criterion;
std::vector<Smooth> smooth;
};
/**
Extension of stpm2 and pstpm2 to include shared frailties with a cluster variable
**/
template<class Stpm2>
class stpm2Shared : public Stpm2 {
public:
stpm2Shared(SEXP sexp) : Stpm2(sexp) {
List list = as<List>(sexp);
IntegerVector cluster = as<IntegerVector>(list["cluster"]);
// wragged array indexed by a map of vectors
for (int i=0; i<cluster.size(); ++i) {
clusters[cluster[i]].push_back(i);
}
}
std::map<int,std::vector<int> > clusters;
};
/**
Extension to the BFGS class
optim() and calc_hessian() assume that data->parscale is an array/vector/NumericVector
optimWithConstraint() further assumes that data->kappa is a double with an initial value
*/
template<class Data>
class BFGS2 : public BFGS {
public:
void optim(optimfn fn, optimgr gr, NumericVector init, void * ex,
bool apply_parscale = true) {
Data * data = (Data *) ex;
n = init.size();
if (apply_parscale) for (int i = 0; i<n; ++i) init[i] /= data->parscale[i];
BFGS::optim(fn,gr,init,ex);
if (apply_parscale) for (int i = 0; i<n; ++i) coef[i] *= data->parscale[i];
hessian = calc_hessian(gr, ex);
}
void optimWithConstraint(optimfn fn, optimgr gr, NumericVector init, void * ex, constraintfn constraint,
bool apply_parscale = true) {
Data * data = (Data *) ex;
n = init.size();
if (apply_parscale) for (int i = 0; i<n; ++i) init[i] /= data->parscale[i];
bool satisfied;
do {
BFGS::optim(fn,gr,init,ex);
satisfied = constraint(n,&coef[0],ex);
if (!satisfied) data->kappa *= 2.0;
} while ((!satisfied)&& data->kappa < 1.0e3);
if (apply_parscale) for (int i = 0; i<n; ++i) coef[i] *= data->parscale[i];
hessian = calc_hessian(gr, ex);
// Rprintf("kappa=%f\n",data->kappa);
}
NumericMatrix calc_hessian(optimgr gr, void * ex) {
Data * data = (Data *) ex;
vec parscale(n);
for (int i=0; i<n; ++i) {
parscale[i] = data->parscale[i];
data->parscale[i] = 1.0;
}
NumericMatrix hessian = BFGS::calc_hessian(gr,ex);
for (int i=0; i<n; ++i) data->parscale[i] = parscale[i];
return hessian;
}
};
template<class Data>
double fminfn(int n, double * beta, void *ex) {
Data * data = (Data *) ex;
vec vbeta(beta,n);
vbeta = vbeta % data->parscale;
vec eta = data->X * vbeta;
vec etaD = data->XD * vbeta;
vec h = data->h(eta,etaD) + data->bhazard;
vec H = data->H(eta,etaD);
double constraint = data->kappa/2.0 * sum(h % h % (h<0)) +
data->kappa/2.0 * sum((H/data->time) % (H/data->time) % (H<0));
vec eps = h*0.0 + 1.0e-16;
h = max(h,eps);
H = max(H,eps);
double ll = sum(data->wt % data->event % log(h)) - sum(data->wt % H) - constraint;
if (data->delayed == 1) {
vec eta0 = data->X0 * vbeta;
vec etaD0 = data->XD0 * vbeta;
mat H0 = data->H(eta0,etaD0);
ll += sum(data->wt0 % H0);
}
return -ll;
}
/**
Objective function for a Gamma shared frailty
Assumes that weights == 1.
The implementation is incomplete, requiring either the use of optimisation without
gradients or the calculation of the gradients.
**/
template<class Data>
double fminfnShared(int n, double * beta, void *ex) {
typedef std::vector<int> Index;
typedef std::map<int,Index> IndexMap;
Data * data = (Data *) ex;
vec vbeta(beta,n-1); // theta is the last parameter in beta
double theta = beta[n-1];
vbeta = vbeta % data->parscale;
vec eta = data->X * vbeta;
vec etaD = data->XD * vbeta;
vec h = data->h(eta,etaD) + data->bhazard;
vec H = data->H(eta,etaD) + data->bhazard;
double constraint = data->kappa/2.0 * sum(h % h % (h<0)); // sum(h^2 | h<0)
vec eps = h*0.0 + 1.0e-16;
h = max(h,eps);
mat H0;
//double ll = sum(data->wt % data->event % log(h)) - sum(data->wt % H) - constraint;
if (data->delayed == 1) {
vec eta0 = data->X0 * vbeta;
vec etaD0 = data->XD0 * vbeta;
H0 = data->H(eta0,etaD0);
//ll += sum(data->wt0 % H0);
}
double ll;
for (IndexMap::iterator it=data->clusters.begin(); it!=data->clusters.end(); ++it) {
int mi=0;
double sumH = 0.0, sumHenter = 0.0;
for (Index::iterator j=it->second.begin(); j!=it->second.end(); ++j) {
if (data->event(*j)==1) {
mi++;
ll += log(h(*j));
}
sumH += H(*j);
if (data->delayed == 1)
sumHenter += H0(*j);
}
ll -= (1.0/theta+mi)*log(1.0+theta*sumH);
if (data->delayed == 1)
ll += 1.0/theta*log(1.0+theta*sumHenter);
if (mi>0) {
int k=1;
for (Index::iterator j=it->second.begin(); j!=it->second.end(); ++j) {
if (data->event(*j)==1) {
ll += log(1.0+theta*(mi-k));
k++;
}
}
}
}
ll -= constraint;
return -ll;
}
/**
Utility function for calculating a gamma frailty log-likelihood within R code
**/
RcppExport SEXP llgammafrailty(SEXP _theta, SEXP _h, SEXP _H, SEXP _d,
SEXP _cluster) {
double theta = as<double>(_theta);
NumericVector h = as<NumericVector>(_h);
NumericVector H = as<NumericVector>(_H);
IntegerVector d = as<IntegerVector>(_d);
IntegerVector cluster = as<IntegerVector>(_cluster);
double ll = 0.0;
// wragged array indexed by a map of vectors
typedef std::vector<int> Index;
typedef std::map<int,Index> IndexMap;
IndexMap clusters;
for (int i=0; i<cluster.size(); ++i) {
clusters[cluster[i]].push_back(i);
}
for (IndexMap::iterator it=clusters.begin(); it!=clusters.end(); ++it) {
int mi=0;
double sumH = 0.0;
for (Index::iterator j=it->second.begin(); j!=it->second.end(); ++j) {
if (d[*j]==1) {
mi++;
ll += log(h[*j]);
}
sumH += H[*j];
// sumHenter += Henter[*j];
}
ll -= (1.0/theta+mi)*log(1.0+theta*sumH);
// ll += 1.0/theta*log(1.0+theta*sumHenter);
if (mi>0) {
int k=1;
for (Index::iterator j=it->second.begin(); j!=it->second.end(); ++j) {
if (d[*j]==1) {
ll += log(1.0+theta*(mi-k));
k++;
}
}
}
}
return wrap(ll);
}
/**
Utility function for calculating a gamma frailty log-likelihood within R code
**/
RcppExport SEXP llgammafrailtydelayed(SEXP _theta, SEXP _h, SEXP _H, SEXP _H0, SEXP _d,
SEXP _cluster) {
double theta = as<double>(_theta);
NumericVector h = as<NumericVector>(_h);
NumericVector H = as<NumericVector>(_H);
NumericVector H0 = as<NumericVector>(_H0); // assumes that _H0 is zero for data not left truncated
IntegerVector d = as<IntegerVector>(_d);
IntegerVector cluster = as<IntegerVector>(_cluster);
double ll = 0.0;
// wragged array indexed by a map of vectors
typedef std::vector<int> Index;
typedef std::map<int,Index> IndexMap;
IndexMap clusters;
for (int i=0; i<cluster.size(); ++i) {
clusters[cluster[i]].push_back(i);
}
for (IndexMap::iterator it=clusters.begin(); it!=clusters.end(); ++it) {
int mi=0;
double sumH = 0.0, sumHenter = 0.0;
for (Index::iterator j=it->second.begin(); j!=it->second.end(); ++j) {
if (d[*j]==1) {
mi++;
ll += log(h[*j]);
}
sumH += H[*j];
sumHenter += H0[*j];
}
ll -= (1.0/theta+mi)*log(1.0+theta*sumH);
ll += 1.0/theta*log(1.0+theta*sumHenter);
if (mi>0) {
int k=1;
for (Index::iterator j=it->second.begin(); j!=it->second.end(); ++j) {
if (d[*j]==1) {
ll += log(1.0+theta*(mi-k));
k++;
}
}
}
}
return wrap(ll);
}
struct CureModel {
int n0, n1, n2;
mat Xshape, Xscale, Xcure;
vec time, status;
};
double fminfn_cureModel(int n, double * beta, void *ex) {
double ll = 0.0;
CureModel * data = (CureModel *) ex;
vec vbeta(beta,n);
vec shape = exp(data->Xshape * vbeta(span(0,data->n0-1)));
vec scale = exp(data->Xscale * vbeta(span(data->n0,data->n1-1)));
vec cure = 1.0/(1.0+exp(-data->Xcure * vbeta(span(data->n1,data->n2-1))));
for (size_t i=0; i < data->time.size(); ++i) {
ll += data->status(i)==1.0 ?
log(1.0-cure(i)) + R::dweibull(data->time(i),shape(i),scale(i),1) :
log(cure(i)+(1.0-cure(i)) * R::pweibull(data->time(i),shape(i),scale(i),0,0));
}
R_CheckUserInterrupt(); /* be polite -- did the user hit ctrl-C? */
return -ll;
}
RcppExport SEXP fitCureModel(SEXP stime, SEXP sstatus, SEXP sXshape,
SEXP sXscale, SEXP sXcure, SEXP sbeta) {
mat Xshape = as<mat>(sXshape);
mat Xscale = as<mat>(sXscale);
mat Xcure = as<mat>(sXcure);
vec time = as<vec>(stime);
vec status = as<vec>(sstatus);
NumericVector init = as<NumericVector>(sbeta);
int n0=Xshape.n_cols;
int n1=n0+Xscale.n_cols;
int n2=n1+Xcure.n_cols;
CureModel data = {n0,n1,n2,Xshape,Xscale,Xcure,time,status};
NelderMead nm;
nm.reltol = 1.0e-8;
nm.maxit = 1000;
nm.optim(& fminfn_cureModel, init, (void *) &data);
for (int i = 0; i<nm.coef.size(); ++i)
init[i] = nm.coef[i]; // clone
return List::create(_("Fmin")=nm.Fmin,
_("coef")=wrap(init),
_("fail")=nm.fail,
_("hessian")=wrap(nm.hessian));
}
template<class Data>
bool fminfn_constraint(int n, double * beta, void *ex) {
Data * data = (Data *) ex;
vec vbeta(beta,n);
vbeta = vbeta % data->parscale;
vec eta = data->X * vbeta;
vec etaD = data->XD * vbeta;
vec h = data->h(eta, etaD) + data->bhazard;
vec H = data->H(eta, etaD);
return all((h>0) % (H>0));
}
template<class Data>
void fminfn_nlm(int n, double * beta, double * f, void *ex) {
*f = fminfn<Data>(n, beta, ex);
};
// log H penalty - using penalty matrices from mgcv
template<class Stpm2>
double pfminfn_SmoothLogH(int n, double * beta, void *ex) {
typedef SmoothLogH Smooth;
typedef pstpm2<Smooth,Stpm2> Data;
double ll = -fminfn<Data>(n,beta,ex);
Data * data = (Data *) ex;
vec vbeta(beta,n);
vbeta = vbeta % data->parscale;
for (size_t i=0; i < data->smooth.size(); ++i) {
Smooth smoothi = data->smooth[i];
ll -= (data->sp)[i]/2 *
dot(vbeta.subvec(smoothi.first_para,smoothi.last_para),
smoothi.S * vbeta.subvec(smoothi.first_para,smoothi.last_para));
}
return -ll;
}
// hazard penalty
template<class Stpm2>
double pfminfn_SmoothHaz(int n, double * beta, void *ex) {
typedef SmoothHaz Smooth;
typedef pstpm2<Smooth,Stpm2> Data;
double ll = -fminfn<Data>(n,beta,ex);
Data * data = (Data *) ex;
vec vbeta(beta,n);
vbeta = vbeta % data->parscale;
for (size_t i=0; i < data->smooth.size(); ++i) {
Smooth obj = data->smooth[i];
vec s0 = obj.X0 * vbeta;
vec s1 = obj.X1 * vbeta;
vec s2 = obj.X2 * vbeta;
vec s3 = obj.X3 * vbeta;
vec h2 = exp(s0) % (s3 + 3*s1%s2 + s1%s1%s1);
ll -= (data->sp)[i]/2 * obj.lambda * sum(obj.w % h2 % h2);
}
return -ll;
};
template<class Data>
void grfn(int n, double * beta, double * gr, void *ex) {
Data * data = (Data *) ex;
vec vbeta(beta,n);
// if (data->trace) Rprint(vbeta);
vbeta = vbeta % data->parscale;
vec eta = data->X * vbeta;
vec etaD = data->XD * vbeta;
vec h = data->h(eta,etaD) + data->bhazard;
vec H = data->H(eta,etaD);
vec one = ones(h.size());
vec eps = h*0.0 + 1.0e-16;
mat gradH = data->gradH(eta,etaD,data->X,data->XD);
mat gradh = data->gradh(eta,etaD,data->X,data->XD);
mat Xgrad = -rmult(gradH, one % (H>eps)) + rmult(gradh, data->event / h % (h>eps));
mat Xconstraint = data->kappa * rmult(gradh,h%(h<eps)) +
data->kappa * rmult(gradH,H / data->time / data->time % (H<eps));
rowvec dconstraint = sum(Xconstraint,0);
// if (data->trace) Rprint(dconstraint);
Xgrad = rmult(Xgrad,data->wt);
rowvec vgr = sum(Xgrad,0);
if (data->delayed == 1) {
vec eta0 = data->X0 * vbeta;
vec etaD0 = data->XD0 * vbeta;
mat gradH0 = data->gradH(eta0, etaD0, data->X0, data->XD0);
vgr += sum(rmult(gradH0, data->wt0),0);
}
for (int i = 0; i<n; ++i) {
gr[i] = -vgr[i]*data->parscale[i] + dconstraint[i];
}
}
template<class Stpm2>
void pgrfn_SmoothLogH(int n, double * beta, double * gr, void *ex) {
typedef SmoothLogH Smooth;
typedef pstpm2<Smooth,Stpm2> Data;
grfn<Data>(n, beta, gr, ex);
Data * data = (Data *) ex;
vec vbeta(beta,n);
vbeta = vbeta % data->parscale;
vec eta = data->X * vbeta;
vec etaD = data->XD * vbeta;
vec h = data->h(eta,etaD) + data->bhazard;
vec H = data->H(eta,etaD);
mat gradh = data->gradh(eta,etaD,data->X,data->XD);
mat gradH = data->gradH(eta,etaD,data->X,data->XD);
rowvec vgr(n, fill::zeros);
for (size_t i=0; i < data->smooth.size(); ++i) {
SmoothLogH smoothi = data->smooth[i];
vgr.subvec(smoothi.first_para,smoothi.last_para) +=
(data->sp)[i] * (smoothi.S * vbeta.subvec(smoothi.first_para,smoothi.last_para)).t();
}
for (int i = 0; i<n; ++i) gr[i] += vgr[i]*data->parscale[i];
}
template<class Stpm2>
void pgrfn_SmoothHaz(int n, double * beta, double * gr, void *ex) {
typedef SmoothHaz Smooth;
typedef pstpm2<Smooth,Stpm2> Data;
grfn<Data>(n, beta, gr, ex);
Data * data = (Data *) ex;
vec vbeta(beta,n);
vbeta = vbeta % data->parscale;
rowvec vgr(n, fill::zeros);
for (size_t i=0; i < data->smooth.size(); ++i) {
Smooth obj = data->smooth[i];
vec s0 = obj.X0 * vbeta;
vec s1 = obj.X1 * vbeta;
vec s2 = obj.X2 * vbeta;
vec s3 = obj.X3 * vbeta;
vec h2 = exp(s0) % (s3 + 3*s1%s2 + s1%s1%s1);
mat dh2sq_dbeta =
2*lmult(h2 % exp(s0),
obj.X3+3*(rmult(obj.X1,s2)+rmult(obj.X2,s1))+3*lmult(s1%s1,obj.X1))+
2*lmult(h2 % h2,obj.X0);
vgr += (data->sp)[i]*obj.lambda*sum(lmult(obj.w, dh2sq_dbeta),0);
}
for (int i = 0; i<n; ++i) gr[i] += vgr[i] * data->parscale[i];
}
// oh, for partial template specialisation for functions...
template<class Smooth,class Stpm2>
double pfminfn(int n, double * beta, void *ex) { return 0.0; }
template<>
double pfminfn<SmoothLogH,stpm2>(int n, double * beta, void *ex) {
return pfminfn_SmoothLogH<stpm2>(n, beta, ex); }
template<>
double pfminfn<SmoothLogH,stpm2PO>(int n, double * beta, void *ex) {
return pfminfn_SmoothLogH<stpm2PO>(n, beta, ex); }
template<>
double pfminfn<SmoothLogH,stpm2Probit>(int n, double * beta, void *ex) {
return pfminfn_SmoothLogH<stpm2Probit>(n, beta, ex); }
template<>
double pfminfn<SmoothLogH,stpm2AH>(int n, double * beta, void *ex) {
return pfminfn_SmoothLogH<stpm2AH>(n, beta, ex); }
template<>
double pfminfn<SmoothHaz,stpm2>(int n, double * beta, void *ex) {
return pfminfn_SmoothHaz<stpm2>(n, beta, ex); }
template<>
double pfminfn<SmoothHaz,stpm2PO>(int n, double * beta, void *ex) {
return pfminfn_SmoothHaz<stpm2PO>(n, beta, ex); }
template<>
double pfminfn<SmoothHaz,stpm2Probit>(int n, double * beta, void *ex) {
return pfminfn_SmoothHaz<stpm2Probit>(n, beta, ex); }
template<class Smooth, class Stpm2>
void pgrfn(int n, double * beta, double * gr, void *ex) { }
template<>
void pgrfn<SmoothLogH,stpm2>(int n, double * beta, double * gr, void *ex) {
pgrfn_SmoothLogH<stpm2>(n, beta, gr, ex); }
template<>
void pgrfn<SmoothLogH,stpm2PO>(int n, double * beta, double * gr, void *ex) {
pgrfn_SmoothLogH<stpm2PO>(n, beta, gr, ex); }
template<>
void pgrfn<SmoothLogH,stpm2Probit>(int n, double * beta, double * gr, void *ex) {
pgrfn_SmoothLogH<stpm2Probit>(n, beta, gr, ex);}
template<>
void pgrfn<SmoothLogH,stpm2AH>(int n, double * beta, double * gr, void *ex) {
pgrfn_SmoothLogH<stpm2AH>(n, beta, gr, ex); }
template<>
void pgrfn<SmoothHaz,stpm2>(int n, double * beta, double * gr, void *ex) {
pgrfn_SmoothHaz<stpm2>(n, beta, gr, ex); }
template<>
void pgrfn<SmoothHaz,stpm2PO>(int n, double * beta, double * gr, void *ex) {
pgrfn_SmoothHaz<stpm2PO>(n, beta, gr, ex); }
template<>
void pgrfn<SmoothHaz,stpm2Probit>(int n, double * beta, double * gr, void *ex) {
pgrfn_SmoothHaz<stpm2Probit>(n, beta, gr, ex); }
template<class Stpm2>
SEXP optim_stpm2(SEXP args) {
Stpm2 data(args);
BFGS2<Stpm2> bfgs;
bfgs.reltol = data.reltol;
bfgs.optimWithConstraint(fminfn<Stpm2>, grfn<Stpm2>, data.init, (void *) &data, fminfn_constraint<Stpm2>);
return List::create(_("fail")=bfgs.fail,
_("coef")=wrap(bfgs.coef),
_("hessian")=wrap(bfgs.hessian));
}
RcppExport SEXP optim_stpm2_ph(SEXP args) {
return optim_stpm2<stpm2>(args); }
RcppExport SEXP optim_stpm2_po(SEXP args) {
return optim_stpm2<stpm2PO>(args); }
RcppExport SEXP optim_stpm2_probit(SEXP args) {
return optim_stpm2<stpm2Probit>(args); }
RcppExport SEXP optim_stpm2_ah(SEXP args) {
return optim_stpm2<stpm2AH>(args); }
RcppExport SEXP optim_stpm2_nlm(SEXP args) {
stpm2 data(args);
data.init = ew_div(data.init,data.parscale);
int n = data.init.size();
Nlm nlm;
nlm.gradtl = nlm.steptl = data.reltol;
//nlm.method=2; nlm.stepmx=0.0;
bool satisfied;
do {
nlm.optim(& fminfn_nlm<stpm2>, & grfn<stpm2>, data.init, (void *) &data);
satisfied = fminfn_constraint<stpm2>(n,&nlm.coef[0],(void *) &data);
if (!satisfied) data.kappa *= 2.0;
} while (!satisfied && data.kappa<1.0e5);
nlm.coef = ew_mult(nlm.coef, data.parscale);
return List::create(_("itrmcd")=wrap(nlm.itrmcd),
_("niter")=wrap(nlm.itncnt),
_("coef")=wrap(nlm.coef),
_("hessian")=wrap(nlm.hessian));
}
template<class Smooth, class Stpm2>
double pstpm2_step_first(double logsp, void * ex) {
typedef pstpm2<Smooth,Stpm2> Data;
Data * data = (Data *) ex;
data->sp[0] = exp(logsp);
BFGS2<Data> bfgs;
bfgs.reltol = data->reltol;
bfgs.optimWithConstraint(pfminfn<Smooth,Stpm2>, pgrfn<Smooth,Stpm2>, data->init, ex, fminfn_constraint<Data>, false); // do not apply parscale b4/after
NumericMatrix hessian0 = bfgs.calc_hessian(grfn<Data>, ex);
if (data->trace > 0) {
Rprintf("Debug on trace calculation. Coef:\n");
Rprint(bfgs.coef);
Rprintf("Hessian0:\n");
Rprint(hessian0);
Rprintf("Hessian:\n");
Rprint(bfgs.hessian);
}
double edf = trace(solve(as<mat>(bfgs.hessian),as<mat>(hessian0)));
double negll = bfgs.calc_objective(fminfn<Data>,ex);
double gcv = negll + data->alpha*edf;
double bic = negll + data->alpha*edf*log(sum(data->event));
data->init = bfgs.coef;
if (data->trace > 0)
Rprintf("sp=%f\tedf=%f\tnegll=%f\tgcv=%f\tbic=%f\talpha=%f\n",data->sp[0],edf,negll,gcv,bic,data->alpha);
return data->criterion==1 ? gcv : bic;
}
template<class Smooth, class Stpm2>
double pstpm2_step_multivariate(int n, double * logsp, void * ex) {
typedef pstpm2<Smooth,Stpm2> Data;
Data * data = (Data *) ex;
double lsp = -9.0, usp = 9.0;
for (int i=0; i < data->sp.size(); ++i)
//data->sp[i] = exp(logsp[i]);
data->sp[i] = exp(bound(logsp[i],lsp,usp));
if (data->trace > 0) {
for (int i = 0; i < data->sp.size(); ++i)
Rprintf("sp[%i]=%f\t",i,data->sp[i]);
}
BFGS2<Data> bfgs;
bfgs.reltol = data->reltol_search;
bfgs.optimWithConstraint(pfminfn<Smooth,Stpm2>, pgrfn<Smooth,Stpm2>, data->init, ex, fminfn_constraint<Data>, false); // do not apply parscale b4/after
NumericMatrix hessian0 = bfgs.calc_hessian(grfn<Data>, ex);
double edf = trace(solve(as<mat>(bfgs.hessian),as<mat>(hessian0)));
double negll = bfgs.calc_objective(fminfn<Data>, ex);
double gcv = negll + data->alpha*edf;
double bic = 2.0*negll + data->alpha*edf*log(sum(data->event));
data->init = bfgs.coef;
// simple boundary constraints
double constraint = 0.0;
// double kappa = 1.0;
for (int i=0; i < data->sp.size(); ++i) {
if (logsp[i] < lsp)
constraint += pow(logsp[i]-lsp,2.0);
if (logsp[i] > usp)
constraint += pow(logsp[i]-usp,2.0);
}
double objective = data->criterion==1 ?
gcv + data->kappa/2.0*constraint :
bic + data->kappa/2.0*constraint;
if (data->trace > 0) {
Rprintf("edf=%f\tnegll=%f\tgcv=%f\tbic=%f\tobjective=%f\n",edf,negll,gcv,bic,objective);
}
return objective;
}
template<class Smooth, class Stpm2>
void pstpm2_step_multivariate_nlm(int n, double * logsp, double * f, void *ex) {
*f = pstpm2_step_multivariate<Smooth,Stpm2>(n, logsp, ex);
};
template<class Smooth, class Stpm2>
SEXP optim_pstpm2_first(SEXP args) {
typedef pstpm2<Smooth,Stpm2> Data;
Data data(args);
int n = data.init.size();
for (int i=0; i<n; ++i) data.init[i] /= data.parscale[i];
double opt_sp = exp(Brent_fmin(log(0.001),log(1000.0),&pstpm2_step_first<Smooth,Stpm2>,&data,1.0e-2));
data.sp[0] = opt_sp;
for (int i=0; i<n; ++i) data.init[i] *= data.parscale[i];
BFGS2<Data> bfgs;
bfgs.coef = data.init;
bfgs.reltol = data.reltol;
bfgs.optimWithConstraint(pfminfn<Smooth,Stpm2>, pgrfn<Smooth,Stpm2>, data.init, (void *) &data, fminfn_constraint<Data>);
NumericMatrix hessian0 = bfgs.calc_hessian(grfn<Data>, (void *) &data);
double edf = trace(solve(as<mat>(bfgs.hessian),as<mat>(hessian0)));
return List::create(_("sp")=wrap(opt_sp),
_("coef")=wrap(bfgs.coef),
_("hessian")=wrap(bfgs.hessian),
_("edf")=wrap(edf));
}
RcppExport SEXP optim_pstpm2LogH_first_ph(SEXP args) {
return optim_pstpm2_first<SmoothLogH,stpm2>(args); }
RcppExport SEXP optim_pstpm2Haz_first_ph(SEXP args) {
return optim_pstpm2_first<SmoothHaz,stpm2>(args); }
RcppExport SEXP optim_pstpm2LogH_first_po(SEXP args) {
return optim_pstpm2_first<SmoothLogH,stpm2PO>(args); }
RcppExport SEXP optim_pstpm2Haz_first_po(SEXP args) {
return optim_pstpm2_first<SmoothHaz,stpm2PO>(args); }
RcppExport SEXP optim_pstpm2LogH_first_probit(SEXP args) {
return optim_pstpm2_first<SmoothLogH,stpm2Probit>(args); }
RcppExport SEXP optim_pstpm2Haz_first_probit(SEXP args) {
return optim_pstpm2_first<SmoothHaz,stpm2Probit>(args); }
RcppExport SEXP optim_pstpm2LogH_first_ah(SEXP args) {
return optim_pstpm2_first<SmoothLogH,stpm2AH>(args); }
template<class Smooth, class Stpm2>
SEXP optim_pstpm2_multivariate(SEXP args) {
typedef pstpm2<Smooth,Stpm2> Data;
Data data(args);
int n = data.init.size();
data.kappa = 10.0;
NelderMead nm;
nm.reltol = 1.0e-5;
nm.maxit = 500;
nm.hessianp = false;
for (int i=0; i<n; ++i) data.init[i] /= data.parscale[i];
NumericVector logsp(data.sp.size());
for (int i=0; i < data.sp.size(); ++i)
logsp[i] = log(data.sp[i]);
bool satisfied;
do {
nm.optim(pstpm2_step_multivariate<Smooth,Stpm2>, logsp, (void *) &data);
satisfied = true;
for (int i=0; i < data.sp.size(); ++i)
if (logsp[i]< -7.0 || logsp[i] > 7.0) satisfied = false;
if (!satisfied) data.kappa *= 2.0;
} while (!satisfied && data.kappa<1.0e5);
for (int i=0; i<n; ++i) data.init[i] *= data.parscale[i];
// for (int i=0; i < data.sp.size(); ++i)
// data.sp[i] = exp(logsp[i]);
for (int i=0; i < nm.coef.size(); ++i)
data.sp[i] = exp(nm.coef[i]);
BFGS2<Data> bfgs;
bfgs.coef = data.init;
bfgs.reltol = data.reltol;
data.kappa = 1.0;
bfgs.optimWithConstraint(pfminfn<Smooth,Stpm2>, pgrfn<Smooth,Stpm2>, data.init, (void *) &data, fminfn_constraint<Data>);
NumericMatrix hessian0 = bfgs.calc_hessian(grfn<Data>, (void *) &data);
double edf = trace(solve(as<mat>(bfgs.hessian),as<mat>(hessian0)));
return List::create(_("sp")=wrap(data.sp),
_("coef")=wrap(bfgs.coef),
_("hessian")=wrap(bfgs.hessian),
_("edf")=wrap(edf));
}
RcppExport SEXP optim_pstpm2LogH_multivariate_ph(SEXP args) {
return optim_pstpm2_multivariate<SmoothLogH,stpm2>(args); }
RcppExport SEXP optim_pstpm2Haz_multivariate_ph(SEXP args) {
return optim_pstpm2_multivariate<SmoothHaz,stpm2>(args); }
RcppExport SEXP optim_pstpm2LogH_multivariate_po(SEXP args) {
return optim_pstpm2_multivariate<SmoothLogH,stpm2PO>(args); }
RcppExport SEXP optim_pstpm2Haz_multivariate_po(SEXP args) {
return optim_pstpm2_multivariate<SmoothHaz,stpm2PO>(args); }
RcppExport SEXP optim_pstpm2LogH_multivariate_probit(SEXP args) {
return optim_pstpm2_multivariate<SmoothLogH,stpm2Probit>(args); }
RcppExport SEXP optim_pstpm2Haz_multivariate_probit(SEXP args) {
return optim_pstpm2_multivariate<SmoothHaz,stpm2Probit>(args); }
RcppExport SEXP optim_pstpm2LogH_multivariate_ah(SEXP args) {
return optim_pstpm2_multivariate<SmoothLogH,stpm2AH>(args); }
template<class Smooth, class Stpm2>
SEXP optim_pstpm2_multivariate_nlm(SEXP args) {
typedef pstpm2<Smooth,Stpm2> Data;
Data data(args);
int n = data.init.size();
data.kappa = 10.0;
Nlm nlm;
nlm.iagflg = 0;
nlm.gradtl = 1.0e-4;
nlm.steptl = 1.0e-4;
nlm.msg = 9 + 4;
nlm.hessianp = false;
for (int i=0; i<n; ++i) data.init[i] /= data.parscale[i];
NumericVector logsp(data.sp.size());
for (int i=0; i < data.sp.size(); ++i)
logsp[i] = log(data.sp[i]);
bool satisfied;
do {
nlm.optim(pstpm2_step_multivariate_nlm<Smooth,Stpm2>, (fcn_p) 0, logsp, (void *) &data);
satisfied = true;
for (int i=0; i < data.sp.size(); ++i)
if (logsp[i]< -7.0 || logsp[i] > 7.0) satisfied = false;
if (!satisfied) data.kappa *= 2.0;
} while (!satisfied && data.kappa<1.0e5);
for (int i=0; i<n; ++i) data.init[i] *= data.parscale[i];
for (int i=0; i < nlm.coef.size(); ++i)
data.sp[i] = exp(nlm.coef[i]);
BFGS2<Data> bfgs;
bfgs.coef = data.init;
bfgs.reltol = data.reltol;
data.kappa = 1.0;
bfgs.optimWithConstraint(pfminfn<Smooth,Stpm2>, pgrfn<Smooth,Stpm2>, data.init, (void *) &data, fminfn_constraint<Data>);
NumericMatrix hessian0 = bfgs.calc_hessian(grfn<Data>, (void *) &data);
double edf = trace(solve(as<mat>(bfgs.hessian),as<mat>(hessian0)));
return List::create(_("sp")=wrap(data.sp),
_("coef")=wrap(bfgs.coef),
_("hessian")=wrap(bfgs.hessian),
_("edf")=wrap(edf));
}
RcppExport SEXP optim_pstpm2LogH_multivariate_ph_nlm(SEXP args) {
return optim_pstpm2_multivariate_nlm<SmoothLogH,stpm2>(args); }
RcppExport SEXP optim_pstpm2Haz_multivariate_ph_nlm(SEXP args) {
return optim_pstpm2_multivariate_nlm<SmoothHaz,stpm2>(args); }
RcppExport SEXP optim_pstpm2LogH_multivariate_po_nlm(SEXP args) {
return optim_pstpm2_multivariate_nlm<SmoothLogH,stpm2PO>(args); }
RcppExport SEXP optim_pstpm2Haz_multivariate_po_nlm(SEXP args) {
return optim_pstpm2_multivariate_nlm<SmoothHaz,stpm2PO>(args); }
RcppExport SEXP optim_pstpm2LogH_multivariate_probit_nlm(SEXP args) {
return optim_pstpm2_multivariate_nlm<SmoothLogH,stpm2Probit>(args); }
RcppExport SEXP optim_pstpm2Haz_multivariate_probit_nlm(SEXP args) {
return optim_pstpm2_multivariate_nlm<SmoothHaz,stpm2Probit>(args); }
template<class Smooth, class Stpm2>
SEXP optim_pstpm2_fixedsp(SEXP args) {
typedef pstpm2<Smooth,Stpm2> Data;
Data data(args);
BFGS2<Data> bfgs;
bfgs.coef = data.init;
bfgs.reltol = data.reltol;
bfgs.optimWithConstraint(pfminfn<Smooth,Stpm2>, pgrfn<Smooth,Stpm2>, data.init, (void *) &data, fminfn_constraint<Data>);
NumericMatrix hessian0 = bfgs.calc_hessian(grfn<Data>, (void *) &data);
double edf = trace(solve(as<mat>(bfgs.hessian),as<mat>(hessian0)));
return List::create(_("sp")=wrap(data.sp),
_("coef")=wrap(bfgs.coef),
_("hessian")=wrap(bfgs.hessian),
_("edf")=wrap(edf));
}
RcppExport SEXP optim_pstpm2LogH_fixedsp_ph(SEXP args) {
return optim_pstpm2_fixedsp<SmoothLogH,stpm2>(args); }
RcppExport SEXP optim_pstpm2Haz_fixedsp_ph(SEXP args) {
return optim_pstpm2_fixedsp<SmoothHaz,stpm2>(args); }
RcppExport SEXP optim_pstpm2LogH_fixedsp_po(SEXP args) {
return optim_pstpm2_fixedsp<SmoothLogH,stpm2PO>(args); }
RcppExport SEXP optim_pstpm2Haz_fixedsp_po(SEXP args) {
return optim_pstpm2_fixedsp<SmoothHaz,stpm2PO>(args); }
RcppExport SEXP optim_pstpm2LogH_fixedsp_probit(SEXP args) {
return optim_pstpm2_fixedsp<SmoothLogH,stpm2Probit>(args); }
RcppExport SEXP optim_pstpm2Haz_fixedsp_probit(SEXP args) {
return optim_pstpm2_fixedsp<SmoothHaz,stpm2Probit>(args); }
RcppExport SEXP optim_pstpm2LogH_fixedsp_ah(SEXP args) {
return optim_pstpm2_fixedsp<SmoothLogH,stpm2AH>(args); }
template<class Smooth, class Stpm2>
SEXP test_pstpm2(SEXP args) {
pstpm2<Smooth,Stpm2> data(args);
int n = data.init.size();
NumericVector init = ew_div(data.init,data.parscale);
double pfmin = pfminfn<Smooth,Stpm2>(n,&init[0],(void *) &data);
NumericVector gr(n);
pgrfn<Smooth,Stpm2>(n,&init[0], &gr[0], (void *) &data);
init = ew_mult(init,data.parscale);
return List::create(_("sp")=wrap(data.sp),
_("pfmin")=wrap(pfmin),
_("gr")=wrap(gr)
);
}
RcppExport SEXP test_pstpm2LogH(SEXP args) {
return test_pstpm2<SmoothLogH,stpm2>(args);
}
RcppExport SEXP test_pstpm2Haz(SEXP args) {
return test_pstpm2<SmoothHaz,stpm2>(args);
}
// R CMD INSTALL ~/src/R/microsimulation
// R -q -e "require(microsimulation); .Call('test_nmmin',1:2,PACKAGE='microsimulation')"
// .Call("optim_stpm2",init,X,XD,rep(bhazard,nrow(X)),wt,ifelse(event,1,0),package="rstpm2")
} // anonymous namespace
