aft_integrated.cpp
#include <RcppArmadillo.h>
#include <c_optim.h>
#include <splines.h>
#include <vector>
namespace rstpm2 {
using namespace arma;
using namespace Rcpp;
// TODO:
// include mixture models in .gradh(), .haz() and .survival()
class aft_integrated {
public:
List args;
vec init;
std::vector<mat> X_vector, X_vector0;
mat Xc;
mat Xt;
vec event;
vec time, time0;
vec gnodes, gweights;
vec boundaryKnots;
vec interiorKnots;
mat q_matrix;
int cure;
int mixture;
ns s;
double kappa; // scale for the quadratic penalty for monotone splines
uvec delayed;
aft_integrated(Rcpp::List args) : args(args) {
init = as<vec>(args("init"));
X_vector = as<std::vector<mat>>(args("X_list"));
X_vector0 = as<std::vector<mat>>(args("X_list0"));
Xt = as<mat>(args("Xt"));
Xc = as<mat>(args("Xc"));
gnodes = as<vec>(args("gnodes"));
gweights = as<vec>(args("gweights"));
event = as<vec>(args("event"));
time = as<vec>(args("time"));
time0 = as<vec>(args("time0"));
boundaryKnots = as<vec>(args("boundaryKnots"));
interiorKnots = as<vec>(args("interiorKnots"));
q_matrix = as<mat>(args("q.const"));
cure = as<int>(args("cure"));
mixture = as<int>(args("mixture"));
s = ns(boundaryKnots, interiorKnots, q_matrix, 1, cure);
kappa = 1.0e3;
delayed = time0>0.0;
}
mat rmult(mat m, vec v) {
mat out(m);
out.each_col() %= v;
return out;
}
mat rmult(mat m, uvec v) {
mat out(m);
out.each_col() %= conv_to<vec>::from(v);
return out;
}
double objective(vec betafull)
{
vec beta, betac, betas, etac;
if (mixture) {
beta = betafull.subvec(0,Xt.n_cols-1);
betac = betafull.subvec(Xt.n_cols, Xt.n_cols + Xc.n_cols - 1);
betas = betafull.subvec(Xt.n_cols+Xc.n_cols, betafull.size()-1);
} else {
beta = betafull.subvec(0,Xt.n_cols-1);
betas = betafull.subvec(Xt.n_cols, betafull.size()-1);
}
vec tstar = time*0.0;
vec scale = time/2.0;
// Can this be done more efficiently using a cube?
for(size_t i=0; i<X_vector.size(); i++) {
tstar += gweights[i]*scale % exp(-X_vector[i]*beta);
}
vec logtstar = log(tstar);
vec etat = Xt*beta;
vec etas = s.basis(logtstar) * betas;
vec etaDs = s.basis(logtstar,1) * betas;
// fix bounds on etaDs
vec eps = etaDs*0. + 1e-8;
double pen = dot(min(etaDs,eps), min(etaDs,eps));
etaDs = max(etaDs, eps);
// add penalty for monotone splines
vec betasStar = s.q_matrix.t() * betas;
for (size_t i=1; i<betasStar.size(); i++) {
double delta = betasStar(i)-betasStar(i-1);
if (delta<0.0)
pen += kappa*delta*delta;
}
vec logh = etas + log(etaDs) - etat - logtstar;
vec H = exp(etas);
double ll = dot(logh,event) - sum(H) - pen; // NB: addition in the gradient
if (mixture) {
etac = Xc * betac;
vec cure_frac = exp(etac)/(1.0+exp(etac));
vec Hu = H, loghu=logh;
vec S = cure_frac + (1-cure_frac) % exp(-Hu);
ll = dot(event, log(1-cure_frac)+loghu-Hu) +
dot(1-event,log(S)) - pen;
}
if (any(delayed)) {
vec tstar0 = time0(delayed)*0.0;
vec scale0 = time0(delayed)/2.0;
for(size_t i=0; i<X_vector0.size(); i++) {
mat X0 = X_vector0[i].rows(delayed);
tstar0 += gweights[i]*scale0 % exp(-X0*beta);
}
vec logtstar0 = log(tstar0);
vec etas0 = s.basis(logtstar0) * betas;
vec etaDs0 = s.basis(logtstar0,1) * betas;
vec H0 = exp(etas0);
vec eps0 = etaDs0*0. + 1e-16;
ll -= dot(min(etaDs0,eps0), min(etaDs0,eps0));
if (!mixture) {
ll += sum(H0);
} else {
vec cure_frac0 = exp(etac(delayed))/(1+exp(etac(delayed)));
vec S0_mix = cure_frac0 + (1.0-cure_frac0) % exp(-H0);
vec H0_mix = -log(S0_mix);
ll += sum(H0_mix);
}
// if (pen>0.0) ll=-100;
}
return -ll;
}
vec gradientPenalty(mat Q, vec beta) { // Q: (nbeta+2) x nbeta
size_t n = Q.n_rows;
mat D = join_rows(zeros(n-1,1),eye(n-1,n-1)) - join_rows(eye(n-1,n-1),zeros(n-1,1)); // (nbeta+1) x (nbeta+2)
vec delta = D * Q * beta; // nbeta+1
mat M = Q.t() * D.row(0).t() * D.row(0) * Q * (delta(0)<0.0); // nbeta x nbeta
for(size_t j=1; j<delta.size(); j++) {
if (delta(j)<0.0)
M += Q.t() * D.row(j).t() * D.row(j) * Q;
}
return 2*M*beta*kappa;
}
vec gradient(vec betafull)
{
vec beta, betac, betas, etac;
if (mixture) {
beta = betafull.subvec(0,Xt.n_cols-1);
betac = betafull.subvec(Xt.n_cols, Xt.n_cols + Xc.n_cols - 1);
betas = betafull.subvec(Xt.n_cols+Xc.n_cols, betafull.size()-1);
} else {
beta = betafull.subvec(0,Xt.n_cols-1);
betas = betafull.subvec(Xt.n_cols, betafull.size()-1);
}
vec tstar = time*0.0;
mat Xtstar = Xt*0.0;
vec scale = time/2.0;
// Can this be done more efficiently using a cube?
for(size_t i=0; i<X_vector.size(); i++) {
vec integrand = gweights[i]*scale % exp(-X_vector[i]*beta);
tstar += integrand;
Xtstar += rmult(X_vector[i],integrand);
}
vec logtstar = log(tstar);
vec etat = Xt*beta;
mat Xs = s.basis(logtstar);
mat XDs = s.basis(logtstar,1);
mat XDDs = s.basis(logtstar,2);
vec etas = Xs * betas;
vec etaDs = XDs * betas;
vec etaDs_old = etaDs;
vec etaDDs = XDDs * betas;
// H calculations
vec H = exp(etas);
mat dHdbeta = -rmult(Xtstar, H % etaDs / tstar);
// penalties
vec eps = etaDs*0. + 1e-8;
uvec pindexs = (etaDs < eps);
// fix bounds on etaDs
mat pgrad = join_rows(2*rmult(Xtstar, etaDs % etaDDs / tstar), -2.0*rmult(XDs,etaDs));
etaDs = max(etaDs, eps);
mat dHdbetas = rmult(Xs,H);
vec logh = etas + log(etaDs) - etat - logtstar;
vec h = exp(logh);
mat dloghdbetas = Xs+rmult(XDs,1/etaDs % (1-pindexs));
mat dloghdbeta = - rmult(Xtstar,etaDDs/etaDs_old/tstar) -
rmult(Xtstar,etaDs_old / tstar) + rmult(Xtstar,1.0/tstar) - Xt;
mat gradi = join_rows(rmult(dloghdbeta,event)-dHdbeta,
rmult(dloghdbetas,event)-dHdbetas) + rmult(pgrad,pindexs);
vec gr = sum(gradi,0).t();
gr -= join_cols(beta*0.0, gradientPenalty(s.q_matrix.t(), betas));
if (mixture) {
etac = Xc * betac;
vec cure_frac = exp(etac)/(1.0+exp(etac));
vec S_mix = cure_frac + (1.0-cure_frac) % exp(-H);
vec H_mix = -log(S_mix);
mat dpidtheta = rmult(Xc, cure_frac % (1-cure_frac));
mat dHdbeta_mix = rmult(dHdbeta,(1.0 - cure_frac) % exp(-H) / S_mix);
mat dHdbetas_mix = rmult(dHdbetas,(1.0 - cure_frac) % exp(-H) / S_mix);
mat dHdtheta_mix = -rmult(dpidtheta, (1.0 - exp(-H))/S_mix);
mat dloghdbeta_mix = dloghdbeta - dHdbeta + dHdbeta_mix;
mat dloghdbetas_mix = dloghdbetas - dHdbetas + dHdbetas_mix;
mat dloghdtheta_mix = dHdtheta_mix - rmult(dpidtheta,1.0 / (1.0 - cure_frac));
mat pgrad = join_rows(2*rmult(Xtstar, etaDs_old % etaDDs / tstar), Xc*0.0, -2.0*rmult(XDs,etaDs_old));
gradi = join_rows(rmult(dloghdbeta_mix,event)-dHdbeta_mix,
rmult(dloghdtheta_mix,event)-dHdtheta_mix,
rmult(dloghdbetas_mix,event)-dHdbetas_mix) + rmult(pgrad,pindexs);
gr = sum(gradi,0).t();
gr -= join_cols(beta*0.0, betac*0.0, gradientPenalty(s.q_matrix.t(), betas));
}
if (any(delayed)) {
vec tstar0 = time0(delayed)*0.0;
vec scale0 = time0(delayed)/2.0;
mat Xtstar0 = Xt.rows(delayed)*0.0;
// Can this be done more efficiently using a cube?
for(size_t i=0; i<X_vector0.size(); i++) {
vec integrand = gweights[i]*scale0 % exp(-X_vector0[i].rows(delayed)*beta);
tstar0 += integrand;
Xtstar0 += rmult(X_vector0[i].rows(delayed),integrand);
}
vec logtstar0 = log(tstar0);
mat Xs0 = s.basis(logtstar0);
mat XDs0 = s.basis(logtstar0,1);
mat XDDs0 = s.basis(logtstar0,2);
vec etas0 = Xs0 * betas;
vec etaDs0 = XDs0 * betas;
vec etaDs0_old = etaDs0;
vec etaDDs0 = XDDs0 * betas;
vec H0 = exp(etas0);
mat dHdbetas0 = rmult(Xs0,H0);
mat dHdbeta0 = -rmult(Xtstar0,H0 % etaDs0 / tstar0);
vec eps0 = etaDs0*0. + 1e-8;
uvec pindexs0 = (etaDs0 < eps0);
etaDs0 = max(etaDs0, eps0);
mat pgrad0 = join_rows(2*rmult(Xtstar0, etaDs0_old % etaDDs0 / tstar0), Xs*0.0);
mat pgrads0 = join_rows(Xt*0.0, -2.0*rmult(XDs0,etaDs0_old));
if (!mixture) {
gr += sum(join_rows(dHdbeta0, dHdbetas0) + rmult(pgrads0,pindexs0) + rmult(pgrad0,pindexs0), 0).t();
} else {
vec cure_frac0 = exp(etac(delayed))/(1.0+exp(etac(delayed)));
vec S_mix0 = cure_frac0 + (1.0-cure_frac0) % exp(-H0);
vec H_mix0 = -log(S_mix0);
mat dpidtheta0 = rmult(Xc.rows(delayed), cure_frac0 % (1-cure_frac0));
mat dHdbeta_mix0 = rmult(dHdbeta0,(1.0 - cure_frac0) % exp(-H0) / S_mix0);
mat dHdbetas_mix0 = rmult(dHdbetas0,(1.0 - cure_frac0) % exp(-H0) / S_mix0);
mat dHdtheta_mix0 = -rmult(dpidtheta0, (1.0 - exp(-H0))/S_mix0);
mat pgrad0 = join_rows(2*rmult(Xtstar0, etaDs0_old % etaDDs0 / tstar0), Xc*0.0, Xs*0.0);
mat pgrads0 = join_rows(Xt*0.0, Xc*0.0, -2.0*rmult(XDs0,etaDs0_old));
gr += sum(join_rows(dHdbeta_mix0, dHdtheta_mix0, dHdbetas_mix0) +
rmult(pgrads0,pindexs0) + rmult(pgrad0,pindexs0), 0).t();
}
}
return -gr;
}
double objective(NumericVector betafull) {
return objective(as<vec>(wrap(betafull)));
}
NumericVector gradient(NumericVector betafull) {
vec value = gradient(as<vec>(wrap(betafull)));
return as<NumericVector>(wrap(value));
}
// vec survival(vec time, std::vector<mat> X_vector1) {
// vec beta = init.subvec(0,X.n_cols-1);
// vec betas = init.subvec(X.n_cols,init.size()-1);
// vec eta = X * beta;
// vec logtstar = log(time) - eta;
// vec etas = s.basis(logtstar) * betas;
// vec S = exp(-exp(etas));
// return S;
// }
// mat haz(vec time, mat X, mat XD)
// {
// vec beta = init.subvec(0,X.n_cols-1);
// vec betas = init.subvec(X.n_cols,init.size()-1);
// vec eta = X * beta;
// vec etaD = XD * beta;
// vec logtstar = log(time) - eta;
// mat Xs = s.basis(logtstar);
// mat XDs = s.basis(logtstar,1);
// mat XDDs = s.basis(logtstar,2);
// vec etas = Xs * betas;
// vec etaDs = XDs * betas;
// vec etaDDs = XDDs * betas;
// // penalties
// vec eps = etaDs*0. + 1e-8;
// uvec pindexs = (etaDs < eps);
// uvec pindex = ((1.0/time - etaD) < eps);
// // fix bounds on etaDs
// etaDs = max(etaDs, eps);
// // fix bounds on etaD
// etaD = 1/time - max(1/time-etaD, eps);
// vec logh = etas + log(etaDs) + log(1/time -etaD);
// vec h = exp(logh);
// return h;
// }
// mat gradh(vec time, mat X, mat XD)
// {
// vec beta = init.subvec(0,X.n_cols-1);
// vec betas = init.subvec(X.n_cols,init.size()-1);
// vec eta = X * beta;
// vec etaD = XD * beta;
// vec logtstar = log(time) - eta;
// mat Xs = s.basis(logtstar);
// mat XDs = s.basis(logtstar,1);
// mat XDDs = s.basis(logtstar,2);
// vec etas = Xs * betas;
// vec etaDs = XDs * betas;
// vec etaDDs = XDDs * betas;
// // penalties
// vec eps = etaDs*0. + 1e-8;
// uvec pindexs = (etaDs < eps);
// uvec pindex = ((1.0/time - etaD) < eps);
// // fix bounds on etaDs
// etaDs = max(etaDs, eps);
// // fix bounds on etaD
// etaD = 1/time - max(1/time-etaD, eps);
// vec logh = etas + log(etaDs) + log(1/time -etaD);
// vec h = exp(logh);
// mat dloghdbetas = Xs+rmult(XDs,1/etaDs % (1-pindexs));
// mat dloghdbeta = -rmult(X,etaDs % (1-pindexs) % (1-pindex)) - rmult(X,etaDDs/etaDs % (1-pindexs) % (1-pindex)) - rmult(XD, (1-pindexs) % (1-pindex)/(1/time-etaD));
// mat gradh = join_rows(rmult(dloghdbeta,h), rmult(dloghdbetas,h));
// return gradh;
// }
};
RcppExport SEXP aft_integrated_model_output(SEXP args) {
using namespace Rcpp;
using namespace arma;
aft_integrated model(args);
List list = as<List>(args);
std::string return_type = as<std::string>(list["return_type"]);
if (return_type == "nmmin") {
// model.pre_process();
NelderMead nm;
nm.trace = as<int>(list["trace"]);
nm.maxit = as<int>(list["maxit"]);
NumericVector betafull = as<NumericVector>(wrap(model.init));
nm.optim<aft_integrated>(betafull,model);
// model.post_process();
return List::create(_("fail")=nm.fail,
_("coef")=wrap(nm.coef),
_("hessian")=wrap(nm.hessian));
}
else if (return_type == "vmmin") {
// model.pre_process();
BFGS bfgs;
bfgs.trace = as<int>(list["trace"]);
bfgs.maxit = as<int>(list["maxit"]);
NumericVector betafull = as<NumericVector>(wrap(model.init));
bfgs.optim<aft_integrated>(betafull,model);
// model.post_process();
return List::create(_("fail")=bfgs.fail,
_("coef")=wrap(bfgs.coef),
_("hessian")=wrap(bfgs.hessian));
}
else if (return_type == "objective")
return wrap(model.objective(model.init));
else if (return_type == "gradient")
return wrap(model.gradient(model.init));
// else if (return_type == "survival")
// return wrap(model.survival(as<vec>(list["time"]),as<mat>(list["X"])));
// else if (return_type == "haz")
// return wrap(model.haz(as<vec>(list["time"]),as<mat>(list["X"]),as<mat>(list["XD"])));
// else if (return_type == "gradh")
// return wrap(model.gradh(as<vec>(list["time"]),as<mat>(list["X"]),as<mat>(list["XD"])));
else {
REprintf("Unknown return_type.\n");
return wrap(-1);
}
}
} // namespace rstpm2
