The R Project SVN
R-packages
(root)/trunk/Old/survival/Examples/test/testall.save – Rev 7500
Rev 104 | Blame | Compare with Previous | Last modification | View Log | Download | RSS feed
S-PLUS : Copyright (c) 1988, 1996 MathSoft, Inc.S : Copyright AT&T.Version 3.4 Release 1 for Sun SPARC, SunOS 5.3 : 1996Mayo local startup...loading survivalloading dateloading slocal.miscset missing value action to 'na.omit'set contrasts to ('contr.treatment', 'contr.poly')stop automatic character -> factor conversion in data framesWorking data will be in .Data> #> # Set up for the test> #> dyn.load("../loadmod.o")> attach("../.Data")> options(na.action="na.omit", contrasts="contr.treatment")> #> # Compute some answers "by hand"> # For test1 and test2 data> byhand1 <- function(coef, newx) {+ s <- exp(coef);+ loglik <- 2*coef - (log(3*s+3) + 2*log(s+3));+ u <- (6 + 3*s - s^2) / ((s+1)*(s+3));+ imat <- s/(s+1)^2 + 6*s/(s+3)^2;++ x <- c(1,1,1,0,0,0);+ status <- c(1,0,1,1,0,1);+ xbar <- c(s/(s+1), s/(s+3), 0, 0);+ haz <- c(1/(3*s+3), 2/(s+3), 0, 1 );+ ties <- c(1,1,2,2,3,4);+ wt <- c(s,s,s,1,1,1);+ mart <- c(1,0,1,1,0,1) - wt* (cumsum(haz))[ties];++ score <- rep(6,0);+ for (i in 1:6) {+ j <- ties[i];+ score[i] <- -wt[i]*(cumsum((x[i]-xbar) * haz))[j];+ score[i] <- score[i] + wt[i]* ((x[i]-xbar)*status[i])[j];+ };++ scho <- c(1/(s+1), (3-s)/(3+s), 0);++ surv <- exp(-cumsum(haz)* exp(coef*newx));+ varhaz.g <- cumsum(c(1/(3*s+3)^2, 2/(s+3)^2, 0, 1 ));++ varhaz.d <- cumsum((newx-xbar) * haz);++ varhaz <- (varhaz.g + varhaz.d0^2/ imat) * exp(2*coef*newx);++ names(xbar) <- names(haz) <- 1:4;+ names(surv) <- names(varhaz) <- 1:4;+ list(loglik=loglik, u=u, imat=imat, xbar=xbar, haz=haz,+ mart=mart, score=score,+ scho=scho, surv=surv, var=varhaz,+ varhaz.g=varhaz.g, varhaz.d=varhaz.d);+ }> > > byhand2 <- function(coef, newx) {+ s <- exp(coef);++ loglik <- 4*coef - log(s+1) - log(s+2) - 3*log(3*s+2) - 2*log(3*s+1);+ u <- 1/(s+1) + 1/(3*s+1) + 4/(3*s+2) -+ ( s/(s+2) +3*s/(3*s+2) + 3*s/(3*s+1));+ imat <- s/(s+1)^2 + 2*s/(s+2)^2 + 6*s/(3*s+2)^2 ++ 3*s/(3*s+1)^2 + 3*s/(3*s+1)^2 + 12*s/(3*s+2)^2;++ hazard <-c( 1/(s+1), 1/(s+2), 1/(3*s+2), 1/(3*s+1), 1/(3*s+1), 2/(3*s+2) );+ xbar <- c(s/(s+1), s/(s+2), 3*s/(3*s+2), 3*s/(3*s+1), 3*s/(3*s+1),+ 3*s/(3*s+2));++ var.g <- cumsum(hazard*hazard /c(1,1,1,1,1,2));+ var.d <- cumsum( (xbar-newx)*hazard);++ surv <- exp(-cumsum(hazard) * exp(coef*newx));+ varhaz <- (var.g + var.d^2/imat)* exp(2*coef*newx);++ list(loglik=loglik, u=u, imat=imat, hazard=hazard,+ xbar=xbar, surv=surv, varhaz=varhaz, var.g=var.g, var.d=var.d);+ }> > byhand3 <- function(coef) {+ #Hard coded -- what is found in the Agreg.3 comments file+ s <- as.vector(exp(coef)) #kill the names attr;++ imat <- s/(s+1)^2 + 2*s/(s+2)^2 + 6*s/(3*s+2)^2 ++ 3*s/(3*s+1)^2 + 3*s/(3*s+1)^2 + 12*s/(3*s+2)^2;++ hazard <-c( 1/(s+1), 1/(s+2), 1/(3*s+2), 1/(3*s+1), 1/(3*s+1), 2/(3*s+2) );+ xbar <- c(s/(s+1), s/(s+2), 3*s/(3*s+2), 3*s/(3*s+1), 3*s/(3*s+1),+ 3*s/(3*s+2));+++ newx <- c(0,0,1,1,1,0,0, 2,2,2,2);+ wt <- exp(coef*newx);+ indx <- c(1,2,4,5,6,1,2,3,4,5,6);++ var.g <- hazard*hazard /c(1,1,1,1,1,2);+ surv <- exp(-cumsum(hazard[indx]*wt));+ var1 <- cumsum(var.g[indx]*wt*wt);+ d <- cumsum( (xbar[indx] - newx)* hazard[indx] * wt);+ var2 <- d^2/imat;+ names(surv) <- names(var1) <-names(var2) <- NULL;+ list(time= c(2,3,7,8,9,12,13,16,17,18,19),+ surv=surv, std= sqrt(var1 + var2), var.g=var1, var.d=var2, d=d,+ hazard=hazard, wt=wt);+ }> > # Create the simplest test data set> #> test1 <- data.frame(time= c(4, 3,1,1,2,2,3),+ status=c(1,NA,1,0,1,1,0),+ x= c(0, 2,1,1,1,0,0))> #> # Do the test on the simple data set> #>> fit <-coxph(Surv(time, status) ~x, test1, method="breslow")> fitCall:coxph(formula = Surv(time, status) ~ x, data = test1, method = "breslow")coef exp(coef) se(coef) z px 1.48 4.37 1.26 1.17 0.24Likelihood ratio test=1.48 on 1 df, p=0.224 n=6 (1 observations deleted due to missing)> fit0 <-coxph(Surv(time, status) ~x, test1, iter=0)> fit0$coefx0> coxph(Surv(time, status) ~x, test1, iter=1, method="breslow")$coefx1.6> coxph(Surv(time, status) ~x, test1, iter=2, method="breslow")$coefx1.472724Warning messages:Ran out of iterations and did not converge in: fitter(X, Y, strats, offset,init = init, iter.max = iter.max, eps = eps, ....> coxph(Surv(time, status) ~x, test1, iter=3, method="breslow")$coefx1.475284> > coxph(Surv(time, status) ~ x, test1, method="efron")Call:coxph(formula = Surv(time, status) ~ x, data = test1, method = "efron")coef exp(coef) se(coef) z px 1.68 5.35 1.28 1.31 0.19Likelihood ratio test=1.84 on 1 df, p=0.175 n=6 (1 observations deleted due to missing)> coxph(Surv(time, status) ~ x, test1, method="exact")Call:coxph(formula = Surv(time, status) ~ x, data = test1, method = "exact")coef exp(coef) se(coef) z px 10.2 26980 116 0.0878 0.93Likelihood ratio test=2.77 on 1 df, p=0.0959 n=6 (1 observations deleted due to missing)Warning messages:Loglik converged before variable 1 ; beta may be infinite. in: fitter(X, Y,strats, offset, init = init, iter.max = iter.max, eps = eps, ....> > resid(fit0)1 2 3 4 5 6 7-0.75 NA 0.8333333 -0.1666667 0.4166667 0.4166667 -0.75> resid(coxph(Surv(time, status) ~1, test1))1 2 3 4 5 6 7-0.75 NA 0.8333333 -0.1666667 0.4166667 0.4166667 -0.75> resid(fit0, "scor")1 2 3 4 5 6 70.2013889 NA 0.4166667 -0.08333333 0.3819444 -0.03472222 0.2013889> resid(fit0, "scho")1 2 2 40.5 0.7916667 -0.2083333 0> > resid(fit)1 2 3 4 5 6 7-0.3333335 NA 0.7287136 -0.2712864 -0.4574266 0.6666665 -0.3333335> resid(fit, "scor")1 2 3 4 5 6 70.2113894 NA 0.1356433 -0.05049747 -0.1262435 -0.3816807 0.2113894> resid(fit, "scho")1 2 2 40.1861408 0.4069299 -0.5930701 0> > predict(fit, type="lp")1 2 3 4 5 6 7-0.737642 NA 0.737642 0.737642 0.737642 -0.737642 -0.737642> predict(fit, type="risk")1 2 3 4 5 6 70.4782403 NA 2.090999 2.090999 2.090999 0.4782403 0.4782403> predict(fit, type="expected")1 2 3 4 5 6 71.333334 NA 0.2712864 0.2712864 1.457427 0.3333335 0.3333335> predict(fit, type="terms")1 2 3 4 5 6 7-0.737642 NA 0.737642 0.737642 0.737642 -0.737642 -0.737642> predict(fit, type="lp", se.fit=T)$fit:1 2 3 4 5 6 7-0.737642 NA 0.737642 0.737642 0.737642 -0.737642 -0.737642$se.fit:[1] 0.6278671 NA 0.6278671 0.6278671 0.6278671 0.6278671 0.6278671> predict(fit, type="risk", se.fit=T)$fit:1 2 3 4 5 6 70.4782403 NA 2.090999 2.090999 2.090999 0.4782403 0.4782403$se.fit:1 2 3 4 5 6 70.434201 NA 0.9079139 0.9079139 0.9079139 0.434201 0.434201> predict(fit, type="expected", se.fit=T)$fit:1 2 3 4 5 6 71.333334 NA 0.2712864 0.2712864 1.457427 0.3333335 0.3333335$se.fit:1 2 3 4 5 6 71.154701 NA 0.5208516 0.5208516 1.207239 0.5773504 0.5773504> predict(fit, type="terms", se.fit=T)$fit:1 2 3 4 5 6 7-0.737642 NA 0.737642 0.737642 0.737642 -0.737642 -0.737642$se.fit:1 2 3 4 5 6 70.6278671 NA 0.6278671 0.6278671 0.6278671 0.6278671 0.6278671> > summary(survfit(fit))Call: survfit.coxph(object = fit)time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.8783 0.122 0.66827 12 4 2 0.4981 0.218 0.21125 14 1 1 0.0615 0.150 0.00051 1> summary(survfit(fit, list(x=2)))Call: survfit.coxph(object = fit, newdata = list(x = 2))time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 3.05e-01 6.50e-01 4.72e-03 12 4 2 1.71e-03 1.98e-02 2.33e-13 14 1 1 8.52e-12 5.29e-10 1.22e-64 1> #> # Create the simple data for agreg> #>> test2 <- data.frame(start=c(1, 2, 5, 2, 1, 7, 3, 4, 8, 8),+ stop =c(2, 3, 6, 7, 8, 9, 9, 9,14,17),+ event=c(1, 1, 1, 1, 1, 1, 1, 0, 0, 0),+ x =c(1, 0, 0, 1, 0, 1, 1, 1, 0, 0) )> #> # Do the test on the simple agreg data set> #> fit <-coxph(Surv(start, stop, event)~ x, test2, method="breslow")> fitCall:coxph(formula = Surv(start, stop, event) ~ x, data = test2, method = "breslow")coef exp(coef) se(coef) z px -0.0845 0.919 0.794 -0.106 0.92Likelihood ratio test=0.01 on 1 df, p=0.915 n= 10> fit0 <-coxph(Surv(start, stop, event)~ x, test2, iter=0)> fit0$coefx0> coxph(Surv(start, stop, event)~ x, test2, iter=1, method="breslow")$coefx-0.08507621> coxph(Surv(start, stop, event)~ x, test2, iter=2, method="breslow")$coefx-0.0845261> coxph(Surv(start, stop, event)~ x, test2, iter=3, method="breslow")$coefx-0.0845261> > coxph(Surv(start, stop, event) ~ x, test2, method="efron")Call:coxph(formula = Surv(start, stop, event) ~ x, data = test2, method = "efron")coef exp(coef) se(coef) z px -0.0211 0.979 0.795 -0.0266 0.98Likelihood ratio test=0 on 1 df, p=0.979 n= 10> coxph(Surv(start, stop, event) ~ x, test2, method="exact")Call:coxph(formula = Surv(start, stop, event) ~ x, data = test2, method = "exact")coef exp(coef) se(coef) z px -0.0916 0.912 0.827 -0.111 0.91Likelihood ratio test=0.01 on 1 df, p=0.912 n= 10> > resid(fit0)1 2 3 4 5 6 7 8 9 100.5 0.6666667 0.8 0.2166667 -0.5333333 0.425 -0.025 -1.15 -0.45 -0.45> resid(fit0, "scor")1 2 3 4 5 6 7 8 9 100.25 -0.2222222 -0.48 -0.1147222 0.1061111 0.245 0.1025 -0.41 0.245 0.245> resid(fit0, "scho")2 3 6 7 8 9 90.5 -0.3333333 -0.6 0.25 -0.75 0.45 0.45> > resid(fit)1 2 3 4 5 6 70.521119 0.6574108 0.7897765 0.2473877 -0.6062935 0.3690249 -0.068765798 9 10-1.068766 -0.4204469 -0.4204469> resid(fit, "scor")1 2 3 4 5 6 70.271565 -0.2069671 -0.4577174 -0.09586133 0.1360823 0.1928898 0.046556528 9 10-0.3738904 0.2436713 0.2436713> resid(fit, "scho")2 3 6 7 8 9 90.521119 -0.3148216 -0.5795531 0.2661809 -0.7338191 0.4204469 0.4204469> > resid(coxph(Surv(start, stop, event)~ x, test2, iter=0, init=log(2)), "score")1 2 3 4 5 6 70.1111111 -0.375 -0.65625 -0.2104592 -0.1712727 0.1327948 0.029478468 9 10-0.3177438 0.2048611 0.2048611> > sfit <-survfit(fit)> sfitCall: survfit.coxph(object = fit)n events mean se(mean) median 0.95LCL 0.95UCL2 7 4.77 2.16 3 2 NA> summary(sfit)Call: survfit.coxph(object = fit)time n.risk n.event survival std.err lower 95% CI upper 95% CI2 2 1 0.607 0.303 0.2279 1.0003 3 1 0.437 0.262 0.1347 1.0006 5 1 0.357 0.226 0.1034 1.0007 4 1 0.277 0.188 0.0729 1.0008 4 1 0.214 0.156 0.0514 0.8949 5 2 0.143 0.112 0.0308 0.667> > # make a doubled data set> temp <- rbind(test2, test2)> temp <- data.frame(temp, x2=c(test2$x, test2$x^2),+ ss=c(rep(0, nrow(test2)), rep(1, nrow(test2))))> fitx <- coxph(Surv(start, stop, event) ~ x2 * strata(ss), data=temp,+ method="breslow")> > sfit <- survfit(fitx, c(fitx$means[1], 0) )> sfitCall: survfit.coxph(object = fitx, newdata = c(fitx$means[1], 0))n events mean se(mean) median 0.95LCL 0.95UCLss=0 2 7 4.77 2.16 3 2 NAss=1 2 7 4.77 2.16 3 2 NA> summary(sfit)Call: survfit.coxph(object = fitx, newdata = c(fitx$means[1], 0))ss=0time n.risk n.event survival std.err lower 95% CI upper 95% CI2 2 1 0.607 0.303 0.2279 1.0003 3 1 0.437 0.262 0.1347 1.0006 5 1 0.357 0.226 0.1034 1.0007 4 1 0.277 0.188 0.0729 1.0008 4 1 0.214 0.156 0.0514 0.8949 5 2 0.143 0.112 0.0308 0.667ss=1time n.risk n.event survival std.err lower 95% CI upper 95% CI2 2 1 0.607 0.346 0.19854 13 3 1 0.437 0.321 0.10358 16 5 1 0.357 0.298 0.06946 17 4 1 0.277 0.275 0.03950 18 4 1 0.214 0.248 0.02218 19 5 2 0.143 0.199 0.00946 1> #> # Even though everyone in strata 1 has x2==0, I won't get the same survival> # curve above if survfit is called without forcing predicted x2 to be> # zero-- otherwise I am asking for a different baseline than the> # simple model did. In this particular case coef[2] is nearly zero, so> # the curves are the same, but the variances differ.> #>> # This mimics 'byhand3' and the documentation> fit <- coxph(Surv(start, stop, event) ~x, test2, method="breslow")> tdata <- data.frame( start=c(0,20, 6,0,5),+ stop =c(5,23,10,5,15),+ event=rep(0,5),+ x=c(0,0,1,0,2) )> > temp <- survfit(fit, tdata, individual=T)> temp2 <- byhand3(fit$coef)> all.equal(temp$surv, temp2$surv)[1] T> all.equal(temp2$std, temp$std.err)[1] T> > temp2$time:[1] 2 3 7 8 9 12 13 16 17 18 19$surv:[1] 0.59385568 0.42159583 0.33011489 0.25848416 0.17564514 0.10430786[7] 0.07405126 0.06200600 0.04952367 0.03955414 0.02773285$std:[1] 0.5575022 0.6851479 0.7089453 0.7356099 0.7654270 0.9576269 1.0413781[8] 1.0108872 1.0307723 1.0978065 1.2970159$var.g:[1] 0.2715650 0.3889323 0.4487646 0.5085969 0.5832373 0.8548022 0.9721696[8] 1.0036852 1.0542114 1.1047377 1.1677689$var.d:[1] 0.039243716 0.080495343 0.053838816 0.032524984 0.002641259 0.062246981[7] 0.112298842 0.018207698 0.008280014 0.100441497 0.514481487$d:[1] 0.24955399 0.35740846 0.29229891 0.22718937 0.06474183 0.31429582[7] 0.42215029 0.16998358 -0.11462913 -0.39924183 -0.90357526$hazard:[1] 0.5211190 0.3425892 0.2102235 0.2661809 0.2661809 0.4204469$wt:[1] 1.0000000 1.0000000 0.9189477 0.9189477 0.9189477 1.0000000 1.0000000[8] 0.8444648 0.8444648 0.8444648 0.8444648> #> # Test out the survival curve and variance, in gory detail> #> xdata <- data.frame(rbind(test1,test1), ss = rep(1:2, rep(nrow(test1),2)))> fit <- coxph(Surv(time, status)~x + strata(ss), xdata, method="breslow")> sfit <- survfit(fit, list(x=0)) #type='aalen' is default> > # From the hand worked notes> bb <- as.vector(exp(fit$coef))> realhaz <- cumsum(c(1/(3*bb+3), 2/(bb+3), 1))> all.equal(sfit$surv, exp(-rep(realhaz,2)))[1] T> > realvar <- cumsum(c( (1/(3*bb+3))^2, 2/(bb+3)^2, 1))> dd <- cumsum(c( (bb/(bb+1))*(1/(3*bb+3)),+ (bb/(bb+3))*(2/(bb+3)), 0*1))> realvar <- realvar + dd^2 * fit$var> all.equal(sfit$std, sqrt(rep(realvar,2)))[1] T> summary(sfit)Call: survfit.coxph(object = fit, newdata = list(x = 0))ss=1time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.940 0.0719 0.8089 12 4 2 0.717 0.1974 0.4176 14 1 1 0.264 0.2734 0.0345 1ss=2time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.940 0.0719 0.8089 12 4 2 0.717 0.1974 0.4176 14 1 1 0.264 0.2734 0.0345 1> > # Get the Kalbfleisch-Prentice survival with Greenwood variance> #> sfit <- survfit(fit, list(x=0), type="kalb")> > # second term is the solution to an equation, in the neighborhood of .6> tfun <- function(alpha) (bb/(1-alpha^bb) + 1/(1-alpha) - (bb+3))^2Warning messages:assigning "tfun" masks an object of the same name on database 2> temp <- nlminb(0.6, tfun, lower=.1, upper=.9)> > realkm <- cumprod(c((1- bb/(3*bb+3))^(1/bb), temp$par, 0))> all.equal(sfit$surv, rep(realkm,2))[1] T> > realvar <- cumsum(c( 1/((3*bb+3)*(2*bb+3)), 2/((bb+3)*2), 1))> dd <- cumsum(c( (bb/(bb+1))*(1/(3*bb+3)),+ (bb/(bb+3))*(2/(bb+3)), 0*1))> realvar <- realvar + dd^2 * fit$var> all.equal(sfit$std, sqrt(rep(realvar,2)))[1] T> summary(sfit)Call: survfit.coxph(object = fit, newdata = list(x = 0), type = "kalb")ss=1time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.930 0.0794 0.787 12 4 2 0.556 0.2333 0.244 14 1 1 0.000 0.0000 0.000 0ss=2time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.930 0.0794 0.787 12 4 2 0.556 0.2333 0.244 14 1 1 0.000 0.0000 0.000 0> > > #> # Repeat with the Efron approximation for Cox, Efron estimates> #> fit <- coxph(Surv(time, status)~x + strata(ss), xdata)> sfit <- survfit(fit, list(x=0))> bb <- as.vector(exp(fit$coef))> > realhaz <- cumsum(c(1/(3*bb+3), 1/(bb+3) + 2/(bb+5), 1))> all.equal(sfit$surv, exp(-rep(realhaz,2)))[1] T> > realvar <- cumsum(c( (1/(3*bb+3))^2, 1/(bb+3)^2 + 4/(bb+5)^2, 1))> dd <- cumsum(c( (bb/(bb+1))*(1/(3*bb+3)),+ (bb/(bb+3))*(1/(bb+3))+ (bb/(bb+5))*(2/(bb+5)), 0*1))> realvar <- realvar + dd^2 * fit$var> all.equal(sfit$std, sqrt(rep(realvar,2)))[1] T> summary(sfit)Call: survfit.coxph(object = fit, newdata = list(x = 0))ss=1time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.949 0.0626 0.8337 12 4 2 0.694 0.2130 0.3801 14 1 1 0.255 0.2670 0.0328 1ss=2time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.949 0.0626 0.8337 12 4 2 0.694 0.2130 0.3801 14 1 1 0.255 0.2670 0.0328 1> > > rm(bb, sfit, realhaz, realvar, fit, realkm, xdata)> #> # The AML data, from Miller, "Survival Analysis", page 49.> #> aml <- list(time= c( 9, 13, 13, 18, 23, 28, 31, 34, 45, 48, 161,+ 5, 5, 8, 8, 12, 16, 23, 27, 30, 33, 43, 45),+ status= c( 1,1,0,1,1,0,1,1,0,1,0, 1,1,1,1,1,0,1,1,1,1,1,1),+ x = as.factor(c(rep("Maintained", 11),+ rep("Nonmaintained", 12) )))> > aml <- data.frame(aml)> #> # These results can be found in Miller> #> fit <- coxph(Surv(aml$time, aml$status) ~ aml$x, method="breslow")> fitCall:coxph(formula = Surv(aml$time, aml$status) ~ aml$x, method = "breslow")coef exp(coef) se(coef) z paml$x 0.904 2.47 0.512 1.77 0.078Likelihood ratio test=3.3 on 1 df, p=0.0694 n= 23> resid(fit, type="mart")1 2 3 4 5 6 70.8622552 0.7920096 -0.2079904 0.7481884 0.6565293 -0.3979666 0.45424898 9 10 11 12 13 140.2547496 -1.05401 -0.5540104 -1.55401 0.8784449 0.8784449 0.740069615 16 17 18 19 20 210.7400696 0.5767732 -0.5137362 0.1516277 0.01702279 -0.1489718 -0.564481622 23-1.151851 -1.603405> resid(fit, type="score")1 2 3 4 5 6 7-0.5468557 -0.4925013 0.141064 -0.4799074 -0.4474163 0.2684542 -0.23590838 9 10 11 12 13 14-0.07265521 0.6408269 0.6408269 0.6408269 0.2377681 0.2377681 0.232585415 16 17 18 19 20 210.2325854 0.2038792 -0.1653081 0.04492356 0.007079928 -0.03965181 -0.181184422 23-0.3950761 -0.4721168> resid(fit, type="scho")5 5 8 8 9 12 130.2706694 0.2706694 0.3081233 0.3081233 -0.6423926 0.3360216 -0.633565318 23 23 27 30 31 33-0.6494303 -0.6791933 0.3208067 0.3269755 0.3360216 -0.597099 0.350569734 43 45 48-0.5525726 0.3778338 0.5484461 0> > fit <- survfit(Surv(aml$time, aml$status) ~ aml$x)> fitCall: survfit(formula = Surv(aml$time, aml$status) ~ aml$x)n events mean se(mean) median 0.95LCL 0.95UCLaml$x=Maintained 11 7 52.6 19.83 31 18 NAaml$x=Nonmaintained 12 11 22.7 4.18 23 8 NA> summary(fit)Call: survfit(formula = Surv(aml$time, aml$status) ~ aml$x)aml$x=Maintainedtime n.risk n.event survival std.err lower 95% CI upper 95% CI9 11 1 0.909 0.0867 0.7541 1.00013 10 1 0.818 0.1163 0.6192 1.00018 8 1 0.716 0.1397 0.4884 1.00023 7 1 0.614 0.1526 0.3769 0.99931 5 1 0.491 0.1642 0.2549 0.94634 4 1 0.368 0.1627 0.1549 0.87548 2 1 0.184 0.1535 0.0359 0.944aml$x=Nonmaintainedtime n.risk n.event survival std.err lower 95% CI upper 95% CI5 12 2 0.8333 0.1076 0.6470 1.0008 10 2 0.6667 0.1361 0.4468 0.99512 8 1 0.5833 0.1423 0.3616 0.94123 6 1 0.4861 0.1481 0.2675 0.88327 5 1 0.3889 0.1470 0.1854 0.81630 4 1 0.2917 0.1387 0.1148 0.74133 3 1 0.1944 0.1219 0.0569 0.66443 2 1 0.0972 0.0919 0.0153 0.62045 1 1 0.0000 NA NA NA> survdiff(Surv(aml$time, aml$status)~ aml$x)Call:survdiff(formula = Surv(aml$time, aml$status) ~ aml$x)N Observed Expected (O-E)^2/E (O-E)^2/Vaml$x=Maintained 11 7 10.69 1.27 3.4aml$x=Nonmaintained 12 11 7.31 1.86 3.4Chisq= 3.4 on 1 degrees of freedom, p= 0.0653> > #> # Test out the weighted K-M> #> # First, equal case weights- shouldn't change the survival, but will> # halve the variance> temp2 <-survfit(Surv(aml$time, aml$status), type="kaplan", weight=rep(2,23))> temp <-survfit(Surv(time, status)~1, aml)> temp$surv/temp2$surv[1] 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1> (temp$std.err/temp2$std.err)^2[1] 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2> > # Risk weights-- use a null Cox model> tfit <- coxph(Surv(aml$time, aml$status) ~ offset(log(1:23)))> sfit <- survfit(tfit, type="aalen")> > # Now compute it by hand> # Ties are a challenge> atime <- sort(aml$time)> denom <- rev(cumsum(rev((1:23)[order(aml$time)])))> denom <- denom[match(unique(atime), atime)]> deaths <- tapply(aml$status, aml$time, sum)> chaz <- cumsum(deaths/denom)> all.equal(sfit$surv, as.vector(exp(-chaz[deaths>0])))[1] T> cvar <- cumsum(deaths/denom^2)> all.equal(sfit$std^2, as.vector(cvar[deaths>0]))[1] T> summary(sfit)Call: survfit.coxph.null(object = tfit, type = "aalen")time n.risk n.event survival std.err lower 95% CI upper 95% CI5 23 2 0.993 0.00509 0.983 1.0008 21 2 0.985 0.00750 0.970 1.0009 19 1 0.980 0.00868 0.964 0.99812 18 1 0.976 0.00970 0.957 0.99513 17 1 0.971 0.01075 0.950 0.99318 14 1 0.966 0.01193 0.943 0.99023 13 2 0.955 0.01400 0.928 0.98327 11 1 0.949 0.01518 0.920 0.97930 9 1 0.942 0.01670 0.910 0.97531 8 1 0.934 0.01856 0.898 0.97133 7 1 0.925 0.02042 0.885 0.96534 6 1 0.913 0.02298 0.870 0.96043 5 1 0.901 0.02567 0.852 0.95345 4 1 0.885 0.03021 0.827 0.94648 2 1 0.843 0.04943 0.752 0.946> > # And the Efron result> summary(survfit(tfit))Call: survfit.coxph.null(object = tfit)time n.risk n.event survival std.err lower 95% CI upper 95% CI5 23 2 0.993 0.00521 0.982 1.0008 21 2 0.984 0.00771 0.970 1.0009 19 1 0.980 0.00885 0.963 0.99812 18 1 0.976 0.00986 0.957 0.99513 17 1 0.971 0.01089 0.950 0.99218 14 1 0.966 0.01205 0.942 0.99023 13 2 0.955 0.01424 0.927 0.98327 11 1 0.948 0.01540 0.919 0.97930 9 1 0.941 0.01689 0.909 0.97531 8 1 0.933 0.01873 0.897 0.97033 7 1 0.924 0.02057 0.884 0.96534 6 1 0.913 0.02310 0.869 0.95943 5 1 0.901 0.02577 0.852 0.95345 4 1 0.884 0.03029 0.826 0.94548 2 1 0.843 0.04944 0.751 0.945> > # Lots of ties, so its a good test case> x1 <- coxph(Surv(time, status)~x, aml, method="efron")> x1Call:coxph(formula = Surv(time, status) ~ x, data = aml, method = "efron")coef exp(coef) se(coef) z px 0.916 2.5 0.512 1.79 0.074Likelihood ratio test=3.38 on 1 df, p=0.0658 n= 23> x2 <- coxph(Surv(rep(0,23),time, status) ~x, aml, method="efron")> x1$coef - x2$coefx2.220446e-16> > rm(x1, x2, atime, denom, deaths, chaz,cvar, tfit, sfit, temp, temp2, fit)> #> # Trivial test of stratified residuals> # Make a second strata = replicate of the first, and I should get the> # exact same answers>> temp <- as.matrix(test1)> n <- nrow(temp)> ndead<- sum(test1$status[!is.na(test1$status)])> temp <- data.frame(rbind(temp, temp)) #later releases of S have rbind.data.frame> tstrat <- rep(1:2, c(n,n))> > fit1 <- coxph(Surv(time, status) ~x, test1)> fit2 <- coxph(Surv(time, status) ~x + strata(tstrat), temp)> > all.equal(resid(fit1) , (resid(fit2))[1:n])[1] T> all.equal(resid(fit1, type="score") , (resid(fit2, type="score"))[1:n])[1] T> all.equal(resid(fit1, type="schoe") , (resid(fit2, type="schoe"))[1:ndead])[1] T> > > #AG model> temp <- as.matrix(test2)> n <- nrow(temp)> ndead<- sum(test2$event[!is.na(test2$event)])> temp <- data.frame(rbind(temp, temp))> tstrat <- rep(1:2, c(n,n))> > fit1 <- coxph(Surv(start, stop, event) ~x, test2)> fit2 <- coxph(Surv(start, stop, event) ~x + strata(tstrat), temp)> > all.equal(resid(fit1) , (resid(fit2))[1:n])[1] T> all.equal(resid(fit1, type="score") , (resid(fit2, type="score"))[1:n])[1] T> all.equal(resid(fit1, type="schoe") , (resid(fit2, type="schoe"))[1:ndead])[1] T> #> # A test to exercise the "infinity" check on 2 variables> #> test3 _ data.frame(futime=1:12, fustat=c(1,0,1,0,1,0,0,0,0,0,0,0),+ x1=rep(0:1,6), x2=c(rep(0,6), rep(1,6)))> coxph(Surv(futime, fustat) ~ x1 + x2, test3)Call:coxph(formula = Surv(futime, fustat) ~ x1 + x2, data = test3)coef exp(coef) se(coef) z px1 -11.2 1.36e-05 156 -0.0716 0.94x2 -11.9 6.51e-06 167 -0.0715 0.94Likelihood ratio test=10.2 on 2 df, p=0.00625 n= 12Warning messages:Loglik converged before variable 1 ,2; beta may be infinite. in: fitter(X,Y, strats, offset, init = init, iter.max = iter.max, eps = eps, ....> # Create a "counting process" version of the simplest test data set> #> test1b<- list(start= c(0, 3, 0, 0, 5, 0, 6,14, 0, 0, 10,20,30, 0),+ stop = c(3,10, 10, 5,20, 6,14,20, 30, 10,20,30,40, 10),+ status=c(0, 1, 0, 0, 1, 0, 0, 1, 0, 0, 0, 0, 1, 0),+ x= c(1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, NA),+ id = c(3, 3, 4, 5, 5, 6, 6, 6, 7, 1, 1, 1, 1, 2))> #> # Check out the various residuals under an Efron approximation> #> fit0 <- coxph(Surv(time, status)~ x, test1, iter=0)> fit <- coxph(Surv(time, status) ~x, test1)> fit0b <- coxph(Surv(start, stop, status) ~ x, test1b, iter=0)> fitb <- coxph(Surv(start, stop, status) ~x, test1b)> fitc <- coxph(Surv(time, status) ~ offset(fit$coef*x), test1)> fitd <- coxph(Surv(start, stop, status) ~ offset(fit$coef*x), test1b)> > resid(fit0)1 2 3 4 5 6 7-0.75 NA 0.8333333 -0.1666667 0.4166667 0.4166667 -0.75> resid(fit0b, collapse=test1b$id)1 2 3 4 5 6 7-0.75 NA 0.8333333 -0.1666667 0.4166667 0.4166667 -0.75> resid(fit)1 2 3 4 5 6 7-0.3655447 NA 0.7191709 -0.2808291 -0.4383383 0.7310859 -0.3655447> resid(fitb, collapse=test1b$id)1 2 3 4 5 6 7-0.3655447 NA 0.7191709 -0.2808291 -0.4383383 0.7310859 -0.3655447> resid(fitc)1 2 3 4 5 6 7-0.3655447 NA 0.7191709 -0.2808291 -0.4383383 0.7310859 -0.3655447> resid(fitd, collapse=test1b$id)1 2 3 4 5 6 7-0.3655447 NA 0.7191709 -0.2808291 -0.4383383 0.7310859 -0.3655447> all.equal(resid(fitc), resid(fit))[1] T> > resid(fit0, type="score")1 2 3 4 5 6 70.2013889 NA 0.4166667 -0.08333333 0.3819444 -0.03472222 0.2013889> resid(fit0b, type="score", collapse=test1b$id)1 2 3 4 5 6 70.2013889 NA 0.4166667 -0.08333333 0.3819444 -0.03472222 0.2013889> resid(fit, type="score")1 2 3 4 5 6 70.2208587 NA 0.1132786 -0.04423418 -0.1029191 -0.4078391 0.2208587> resid(fitb, type="score", collapse=test1b$id)1 2 3 4 5 6 70.2208587 NA 0.1132786 -0.04423418 -0.1029191 -0.4078391 0.2208587> > resid(fit0, type="scho")1 2 2 40.5 0.7916667 -0.2083333 0> resid(fit0b, type="scho", collapse=test1b$id)10 20 20 400.5 0.7916667 -0.2083333 0> resid(fit, type="scho")1 2 2 40.1575128 0.4212454 -0.5787546 0> resid(fitb, type="scho", collapse=test1b$id)10 20 20 400.1575128 0.4212454 -0.5787546 0> > summary(survfit(fit0, list(x=0)))Call: survfit.coxph(object = fit0, newdata = list(x = 0))time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.846 0.169 0.5725 12 4 2 0.472 0.246 0.1700 14 1 1 0.174 0.196 0.0191 1> summary(survfit(fit, list(x=0)))Call: survfit.coxph(object = fit, newdata = list(x = 0))time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.949 0.0732 0.8157 12 4 2 0.694 0.2541 0.3385 14 1 1 0.255 0.2718 0.0317 1> summary(survfit(fitb,list(x=0)))Call: survfit.coxph(object = fitb, newdata = list(x = 0))time n.risk n.event survival std.err lower 95% CI upper 95% CI10 6 1 0.949 0.0732 0.8157 120 4 2 0.694 0.2541 0.3385 140 1 1 0.255 0.2718 0.0317 1> summary(survfit(fit))Call: survfit.coxph(object = fit)time n.risk n.event survival std.err lower 95% CI upper 95% CI1 6 1 0.8857 0.117 0.683035 12 4 2 0.4294 0.237 0.145743 14 1 1 0.0425 0.116 0.000198 1> # Tests of the weighted Cox model> #> # Similar data set to test1, but add weights,> # a double-death/censor tied time> # a censored last subject> # The latter two are cases covered only feebly elsewhere.> testw1 <- data.frame(time= c(1,1,2,2,2,2,3,4,5),+ status= c(1,0,1,1,1,0,0,1,0),+ x= c(2,0,1,1,0,1,0,1,0),+ wt = c(1,2,3,4,3,2,1,2,1))> > fit0 <- coxph(Surv(time, status) ~x, testw1, weights=wt,+ method="breslow", iter=0)> fit <- coxph(Surv(time, status) ~x, testw1, weights=wt, method="breslow")> fit0Call:coxph(formula = Surv(time, status) ~ x, data = testw1, weights = wt, iter.max= 0, method = "breslow")coef exp(coef) se(coef) z px 0 1 0.586 0 1Likelihood ratio test=0 on 1 df, p=1 n= 9> summary(fit)Call:coxph(formula = Surv(time, status) ~ x, data = testw1, weights = wt, method ="breslow")n= 9coef exp(coef) se(coef) z px 0.86 2.36 0.713 1.21 0.23exp(coef) exp(-coef) lower .95 upper .95x 2.36 0.423 0.584 9.56Rsquare= 0.171 (max possible= 0.999 )Likelihood ratio test= 1.69 on 1 df, p=0.193Wald test = 1.45 on 1 df, p=0.228Score (logrank) test = 1.52 on 1 df, p=0.217> resid(fit0, type="mart")1 2 3 4 5 6 70.9473684 -0.05263158 0.3223684 0.3223684 0.3223684 -0.6776316 -0.67763168 9-0.3442982 -1.344298> resid(fit0, type="score")1 2 3 4 5 6 71.246537 0.03601108 0.100567 0.100567 -0.2218014 -0.211933 0.46569868 9-0.1008219 0.910143> resid(fit0, type="scho")1 2 2 2 41.315789 0.3125 0.3125 -0.6875 0.3333333> fit0 <- coxph(Surv(time, status) ~x,testw1, weights=wt, iter=0)> resid(fit0, "mart")1 2 3 4 5 6 70.9473684 -0.05263158 0.4445489 0.4445489 0.4445489 -0.8812657 -0.88126578 9-0.5479323 -1.547932> resid(coxph(Surv(time, status) ~1, testw1, weights=wt)) #Null model1 2 3 4 5 6 70.9473684 -0.05263158 0.4445489 0.4445489 0.4445489 -0.8812657 -0.88126578 9-0.5479323 -1.547932> > resid(fit, type="mart")1 2 3 4 5 6 70.8553129 -0.02593186 0.1763631 0.1763631 0.6513114 -0.8236369 -0.34868868 9-0.64894 -0.6980825> resid(fit, type="score")1 2 3 4 5 6 70.8868191 0.02497664 0.03608986 0.03608986 -0.5429742 -0.1252885 0.29564748 9-0.09476947 0.5840033> resid(fit, type="scho")1 2 2 2 41.036836 0.1613784 0.1613784 -0.8386216 0.174697> > fit <- coxph(Surv(time, status) ~x, testw1, weights=wt, method="efron")> fitCall:coxph(formula = Surv(time, status) ~ x, data = testw1, weights = wt, method ="efron")coef exp(coef) se(coef) z px 0.873 2.39 0.713 1.22 0.22Likelihood ratio test=1.75 on 1 df, p=0.186 n= 9> resid(fit, type="mart")1 2 3 4 5 6 70.8533465 -0.02560734 0.3226534 0.3226534 0.7169606 -1.077725 -0.45034368 9-0.9049011 -0.7959915> resid(fit, type="score")1 2 3 4 5 6 70.8811638 0.0247726 0.06057842 0.06057842 -0.5972381 -0.1673716 0.38040488 9-0.1375035 0.6663165> resid(fit, type="scho")1 2 2 2 41.032598 0.1621769 0.1621769 -0.8378231 0.172824> # Tests of the weighted Cox model, AG form of the data> # Same solution as doweight1.s> #> testw2 <- data.frame(id = c( 1, 1, 2, 3, 3, 3, 4, 5, 5, 6, 7, 8, 8, 9),+ begin= c( 0, 5, 0, 0,10,15, 0, 0,14, 0, 0, 0,23, 0),+ time= c( 5,10,10,10,15,20,20,14,20,20,30,23,40,50),+ status= c( 0, 1, 0, 0, 0, 1, 1, 0, 1, 0, 0, 0, 1, 0),+ x= c( 2, 2, 0, 1, 1, 1, 1, 0, 0, 1, 0, 1, 1, 0),+ wt = c( 1, 1, 2, 3, 3, 3, 4, 3, 3, 2, 1, 2, 2, 1))> > fit0 <- coxph(Surv(begin,time, status) ~x, testw2, weights=wt,+ method="breslow", iter=0)> fit <- coxph(Surv(begin,time, status) ~x, testw2, weights=wt, method="breslow")> fit0Call:coxph(formula = Surv(begin, time, status) ~ x, data = testw2, weights = wt,iter.max = 0, method = "breslow")coef exp(coef) se(coef) z px 0 1 0.586 0 1Likelihood ratio test=0 on 1 df, p=1 n= 14> summary(fit)Call:coxph(formula = Surv(begin, time, status) ~ x, data = testw2, weights = wt,method = "breslow")n= 14coef exp(coef) se(coef) z px 0.86 2.36 0.713 1.21 0.23exp(coef) exp(-coef) lower .95 upper .95x 2.36 0.423 0.584 9.56Rsquare= 0.114 (max possible= 0.991 )Likelihood ratio test= 1.69 on 1 df, p=0.193Wald test = 1.45 on 1 df, p=0.228Score (logrank) test = 1.52 on 1 df, p=0.217> resid(fit0, type="mart", collapse=testw2$id)1 2 3 4 5 6 70.9473684 -0.05263158 0.3223684 0.3223684 0.3223684 -0.6776316 -0.67763168 9-0.3442982 -1.344298> resid(fit0, type="score", collapse=testw2$id)1 2 3 4 5 6 71.246537 0.03601108 0.100567 0.100567 -0.2218014 -0.211933 0.46569868 9-0.1008219 0.910143> resid(fit0, type="scho")10 20 20 20 401.315789 0.3125 0.3125 -0.6875 0.3333333> > resid(fit, type="mart", collapse=testw2$id)1 2 3 4 5 6 70.8553129 -0.02593186 0.1763631 0.1763631 0.6513114 -0.8236369 -0.34868868 9-0.64894 -0.6980825> resid(fit, type="score", collapse=testw2$id)1 2 3 4 5 6 70.8868191 0.02497664 0.03608986 0.03608986 -0.5429742 -0.1252885 0.29564748 9-0.09476947 0.5840033> resid(fit, type="scho")10 20 20 20 401.036836 0.1613784 0.1613784 -0.8386216 0.174697> fit0 <- coxph(Surv(begin, time, status) ~x,testw2, weights=wt, iter=0)> resid(fit0, "mart", collapse=testw2$id)1 2 3 4 5 6 70.9473684 -0.05263158 0.4445489 0.4445489 0.4445489 -0.8812657 -0.88126578 9-0.5479323 -1.547932> resid(coxph(Surv(begin, time, status) ~1, testw2, weights=wt)+ , collapse=testw2$id) #Null model1 2 3 4 5 6 70.9473684 -0.05263158 0.4445489 0.4445489 0.4445489 -0.8812657 -0.88126578 9-0.5479323 -1.547932> > fit <- coxph(Surv(begin,time, status) ~x, testw2, weights=wt, method="efron")> fitCall:coxph(formula = Surv(begin, time, status) ~ x, data = testw2, weights = wt,method = "efron")coef exp(coef) se(coef) z px 0.873 2.39 0.713 1.22 0.22Likelihood ratio test=1.75 on 1 df, p=0.186 n= 14> resid(fit, type="mart", collapse=testw2$id)1 2 3 4 5 6 70.8533465 -0.02560734 0.3226534 0.3226534 0.7169606 -1.077725 -0.45034368 9-0.9049011 -0.7959915> resid(fit, type="score", collapse=testw2$id)1 2 3 4 5 6 70.8811638 0.0247726 0.06057842 0.06057842 -0.5972381 -0.1673716 0.38040488 9-0.1375035 0.6663165> resid(fit, type="scho")10 20 20 20 401.032598 0.1621769 0.1621769 -0.8378231 0.172824> #> # The test data set from Turnbull, JASA 1974, 169-73.> #> # status 0=right censored> # 1=exact> # 2=left censored> #>> turnbull <- data.frame( time =c( 1,1,1, 2,2,2, 3,3,3, 4,4,4),+ status=c( 1,0,2, 1,0,2, 1,0,2, 1,0,2),+ n =c(12,3,2, 6,2,4, 2,0,2, 3,3,5))> #> # Compute the K-M for the Turnbull data> # via a slow EM calculation> #>> emsurv <- function(time, status, wt, verbose=T) {+ left.cen <- (status==2);+ if (!any(left.cen)) stop("No left censored data!");+ if (!any(status==1))stop("Must have some exact death times");++ tempy <- Surv(time[!left.cen], status[!left.cen]);+ ww <- wt[!left.cen];+ tempx <- factor(rep(1, sum(!left.cen)));+ tfit <- survfit.km(tempx, tempy, casewt=ww);+ if (verbose)+ cat("Iteration 0, survival=", format(round(tfit$surv[tfit$n.event>0],3)),+ "\n");++ stimes <- tfit$time[tfit$n.event>0];+ ltime <- time[left.cen];+ lwt <- wt[left.cen];+ tempx <- factor(rep(1, length(stimes) + sum(!left.cen)));+ tempy <- Surv(c(time[!left.cen], stimes),+ c(status[!left.cen], rep(1, length(stimes))));+ for (iter in 1:4) {+ wt2 <- stimes*0;+ ssurv <- tfit$surv[tfit$n.event>0];+ sjump <- diff(c(1, ssurv));+ for (j in 1:(length(ltime))) {+ k <- sum(ltime[j]>=stimes) #index of the death time;+ if (k==0)+ stop("Left censored observation before the first death");+ wt2[1:k] <- wt2[1:k] + lwt[j]*sjump[1:k] /(ssurv[k]-1);+ };+ tfit <- survfit.km(tempx, tempy, casewt=c(ww, wt2));+ if (verbose) {+ cat("Iteration", iter, "survival=",+ format(round(tfit$surv[tfit$n.event>0],3)), "\n");+ cat(" weights=", format(round(wt2,3)), "\n");+ };+ };+ survfit(tempy ~ tempx, weights=c(ww, wt2));+ }> > temp <-emsurv(turnbull$time, turnbull$status, turnbull$n)Iteration 0, survival= 0.613 0.383 0.287 0.144Iteration 1 survival= 0.549 0.303 0.214 0.094weights= 7.856 3.477 0.828 0.839Iteration 2 survival= 0.540 0.296 0.210 0.095weights= 8.228 3.394 0.714 0.664Iteration 3 survival= 0.538 0.295 0.210 0.095weights= 8.315 3.356 0.690 0.638Iteration 4 survival= 0.538 0.295 0.210 0.095weights= 8.338 3.342 0.685 0.635> print(summary(temp))Call: survfit(formula = tempy ~ tempx, weights = c(ww, wt2))time n.risk n.event survival std.err lower 95% CI upper 95% CI1 44.00 20.34 0.5378 0.0752 0.4089 0.7072 20.66 9.34 0.2946 0.0719 0.1827 0.4753 9.32 2.68 0.2098 0.0673 0.1119 0.3934 6.64 3.64 0.0948 0.0507 0.0333 0.270> #> # Read in the ovarian data> #>> ovarian <- read.table("data.ovarian", row.names=NULL,+ col.names= c("futime", "fustat", "age", "resid.ds", "rx", "ecog.ps"))> #> # Test the coxph program on the Ovarian data> #>> xx _ order(ovarian$futime) #put data in same order as SAS green book> temp <- ovarian[xx,]> attach(temp)> > # List the data> tempfutime fustat age resid.ds rx ecog.ps1 59 1 72.3315 2 1 12 115 1 74.4932 2 1 13 156 1 66.4658 2 1 222 268 1 74.5041 2 1 223 329 1 43.1370 2 1 124 353 1 63.2192 1 2 225 365 1 64.4247 2 2 126 377 0 58.3096 1 2 14 421 0 53.3644 2 2 15 431 1 50.3397 2 1 16 448 0 56.4301 1 1 27 464 1 56.9370 2 2 28 475 1 59.8548 2 2 29 477 0 64.1753 2 1 110 563 1 55.1781 1 2 211 638 1 56.7562 1 1 212 744 0 50.1096 1 2 113 769 0 59.6301 2 2 214 770 0 57.0521 2 2 115 803 0 39.2712 1 1 116 855 0 43.1233 1 1 217 1040 0 38.8932 2 1 218 1106 0 44.6000 1 1 119 1129 0 53.9068 1 2 120 1206 0 44.2055 2 2 121 1227 0 59.5890 1 2 2> > summary(survfit(Surv(futime, fustat)), censor=T)Call: survfit(formula = Surv(futime, fustat))time n.risk n.event survival std.err lower 95% CI upper 95% CI59 26 1 0.962 0.0377 0.890 1.000115 25 1 0.923 0.0523 0.826 1.000156 24 1 0.885 0.0627 0.770 1.000268 23 1 0.846 0.0708 0.718 0.997329 22 1 0.808 0.0773 0.670 0.974353 21 1 0.769 0.0826 0.623 0.949365 20 1 0.731 0.0870 0.579 0.923377 19 0 0.731 0.0870 0.579 0.923421 18 0 0.731 0.0870 0.579 0.923431 17 1 0.688 0.0919 0.529 0.894448 16 0 0.688 0.0919 0.529 0.894464 15 1 0.642 0.0965 0.478 0.862475 14 1 0.596 0.0999 0.429 0.828477 13 0 0.596 0.0999 0.429 0.828563 12 1 0.546 0.1032 0.377 0.791638 11 1 0.497 0.1051 0.328 0.752744 10 0 0.497 0.1051 0.328 0.752769 9 0 0.497 0.1051 0.328 0.752770 8 0 0.497 0.1051 0.328 0.752803 7 0 0.497 0.1051 0.328 0.752855 6 0 0.497 0.1051 0.328 0.7521040 5 0 0.497 0.1051 0.328 0.7521106 4 0 0.497 0.1051 0.328 0.7521129 3 0 0.497 0.1051 0.328 0.7521206 2 0 0.497 0.1051 0.328 0.7521227 1 0 0.497 0.1051 0.328 0.752> > # Various models> coxph(Surv(futime, fustat)~ age)Call:coxph(formula = Surv(futime, fustat) ~ age)coef exp(coef) se(coef) z page 0.162 1.18 0.0497 3.25 0.0012Likelihood ratio test=14.3 on 1 df, p=0.000156 n= 26> coxph(Surv(futime, fustat)~ resid.ds)Call:coxph(formula = Surv(futime, fustat) ~ resid.ds)coef exp(coef) se(coef) z presid.ds 1.21 3.35 0.672 1.8 0.072Likelihood ratio test=3.76 on 1 df, p=0.0525 n= 26> coxph(Surv(futime, fustat)~ rx)Call:coxph(formula = Surv(futime, fustat) ~ rx)coef exp(coef) se(coef) z prx -0.596 0.551 0.587 -1.02 0.31Likelihood ratio test=1.05 on 1 df, p=0.305 n= 26> coxph(Surv(futime, fustat)~ ecog.ps)Call:coxph(formula = Surv(futime, fustat) ~ ecog.ps)coef exp(coef) se(coef) z pecog.ps 0.398 1.49 0.586 0.679 0.5Likelihood ratio test=0.47 on 1 df, p=0.494 n= 26> > coxph(Surv(futime, fustat)~ resid.ds + rx + ecog.ps)Call:coxph(formula = Surv(futime, fustat) ~ resid.ds + rx + ecog.ps)coef exp(coef) se(coef) z presid.ds 1.347 3.844 0.680 1.980 0.048rx -0.749 0.473 0.595 -1.260 0.210ecog.ps 0.453 1.573 0.590 0.767 0.440Likelihood ratio test=6.03 on 3 df, p=0.11 n= 26> coxph(Surv(futime, fustat)~ age + rx + ecog.ps)Call:coxph(formula = Surv(futime, fustat) ~ age + rx + ecog.ps)coef exp(coef) se(coef) z page 0.147 1.158 0.0463 3.17 0.0015rx -0.815 0.443 0.6342 -1.28 0.2000ecog.ps 0.103 1.109 0.6064 0.17 0.8600Likelihood ratio test=15.9 on 3 df, p=0.00118 n= 26> coxph(Surv(futime, fustat)~ age + resid.ds + ecog.ps)Call:coxph(formula = Surv(futime, fustat) ~ age + resid.ds + ecog.ps)coef exp(coef) se(coef) z page 0.142 1.15 0.052 2.740 0.0061resid.ds 0.663 1.94 0.750 0.883 0.3800ecog.ps 0.166 1.18 0.615 0.271 0.7900Likelihood ratio test=15.1 on 3 df, p=0.00173 n= 26> coxph(Surv(futime, fustat)~ age + resid.ds + rx)Call:coxph(formula = Surv(futime, fustat) ~ age + resid.ds + rx)coef exp(coef) se(coef) z page 0.129 1.137 0.0473 2.718 0.0066resid.ds 0.696 2.006 0.7585 0.918 0.3600rx -0.849 0.428 0.6392 -1.328 0.1800Likelihood ratio test=16.8 on 3 df, p=0.000789 n= 26> > # Residuals> fit <- coxph(Surv(futime, fustat)~ age + resid.ds + rx + ecog.ps )> resid(fit)1 2 3 22 23 24 250.8410328 0.5442439 0.5967083 -0.6548 0.9586613 0.8211167 0.551365526 4 5 6 7 8 9-0.09154016 -0.1128138 0.7511159 -0.3260903 0.5999893 0.2957072 -2.15325810 11 12 13 14 15 160.7624347 0.06474273 -0.1168075 -1.225628 -0.6347484 -0.07535826 -0.170589117 18 19 20 21-0.2298604 -0.1465486 -0.1876292 -0.1277155 -0.5337312> resid(fit, "dev")1 2 3 22 23 24 25 261.412816 0.6950591 0.78916 -0.5497633 2.11059 1.34157 0.7073631 -0.42787894 5 6 7 8 9 10-0.4750027 1.131063 -0.807577 0.7953296 0.3312216 -2.075215 1.1617911 12 13 14 15 16 170.06619521 -0.4833375 -1.565649 -1.12672 -0.3882223 -0.5841046 -0.678027118 19 20 21-0.5413846 -0.6125834 -0.5054018 -1.033181> resid(fit, "scor")age resid.ds rx ecog.ps1 2.26503312 0.05686358 -0.10565381 -0.426616872 3.02525487 0.04641313 -0.08623664 -0.348212753 -0.06851286 0.07131432 -0.13250360 0.0616752922 -3.48057915 -0.03965964 0.07368850 -0.2666932923 -14.86623701 0.28137018 -0.52279209 -0.4388115024 3.96084302 -0.56566918 0.34648947 0.4490740825 4.30025736 0.15241263 0.22417524 -0.2039043826 0.31490645 0.07091765 -0.05212199 0.048456244 0.94597636 -0.02541511 -0.06423497 0.059717305 -5.41507108 0.21605963 -0.32258093 -0.393339096 1.48999549 0.24899474 0.14035144 -0.153806657 -0.68612390 0.13740892 0.28392480 0.291965048 0.93116933 0.08428959 0.16040158 0.184306409 -8.20092606 -0.51356176 0.95647604 1.1133710610 0.95287551 -0.31078221 0.21463990 0.1736338911 2.85526174 0.09417729 -0.14186604 -0.0758608412 0.92721126 0.07495004 -0.05400752 0.0706157913 -1.93962989 -0.43919870 -0.56668532 -0.4846767214 0.63185377 -0.22745951 -0.29348439 0.3837360215 1.41495229 0.04835394 0.04051536 0.0455577016 2.54591219 0.10945918 0.09171494 -0.0674597617 4.40282419 -0.08236954 0.12358138 -0.0908987118 1.97071871 0.09403354 0.07878993 0.0885957219 0.77692382 0.12039307 -0.08675288 0.1134309120 1.76784314 -0.04576633 -0.05905097 0.0772101821 -0.82272543 0.34247079 -0.24677772 -0.21106496> resid(fit, "scho")age resid.ds rx ecog.ps59 2.69315678 0.06761161 -0.1256239 -0.5072536115 5.36390193 0.08039118 -0.1493686 -0.6031317156 -0.89877404 0.10683988 -0.1985109 0.1984379268 6.95664457 0.12857952 -0.2389036 0.2388158329 -15.73656567 0.28889884 -0.5367805 -0.4634169353 4.06104424 -0.70587652 0.4535120 0.5282024365 5.50035871 0.25348266 0.4796229 -0.4413864431 -8.06809462 0.27490178 -0.4297023 -0.5248323464 -2.15471513 0.23158423 0.5066040 0.4814387475 0.57065101 0.25226661 0.5518479 0.5244351563 0.06487254 -0.47274521 0.3319974 0.2747028638 1.64752693 -0.50593435 -0.6446946 0.2939883> > fit <- coxph(Surv(futime, fustat) ~ age + ecog.ps + strata(rx))> summary(fit)Call:coxph(formula = Surv(futime, fustat) ~ age + ecog.ps + strata(rx))n= 26coef exp(coef) se(coef) z page 0.1385 1.149 0.048 2.885 0.0039ecog.ps -0.0967 0.908 0.630 -0.154 0.8800exp(coef) exp(-coef) lower .95 upper .95age 1.149 0.871 1.045 1.26ecog.ps 0.908 1.102 0.264 3.12Rsquare= 0.387 (max possible= 0.874 )Likelihood ratio test= 12.7 on 2 df, p=0.00174Wald test = 8.43 on 2 df, p=0.0148Score (logrank) test = 12.2 on 2 df, p=0.0022> summary(survfit(fit))Call: survfit.coxph(object = fit)rx=1time n.risk n.event survival std.err lower 95% CI upper 95% CI59 13 1 0.978 0.0266 0.9275 1115 12 1 0.951 0.0478 0.8620 1156 11 1 0.910 0.0760 0.7722 1268 10 1 0.861 0.1055 0.6776 1329 9 1 0.736 0.1525 0.4909 1431 8 1 0.627 0.1704 0.3680 1638 5 1 0.333 0.2296 0.0865 1rx=2time n.risk n.event survival std.err lower 95% CI upper 95% CI353 13 1 0.943 0.0560 0.839 1.000365 12 1 0.880 0.0812 0.735 1.000464 9 1 0.789 0.1143 0.594 1.000475 8 1 0.697 0.1350 0.477 1.000563 7 1 0.597 0.1494 0.366 0.975> sfit <- survfit(fit, list(age=c(30,70), ecog.ps=c(2,3))) #two columns> sfitCall: survfit.coxph(object = fit, newdata = list(age = c(30, 70), ecog.ps = c(2, 3)))n events mean se(mean) median 0.95LCL 0.95UCLrx=1 13 7 634 98.9 NA NA NArx=1 13 7 259 24.7 268 115 NArx=2 13 5 562 29.0 NA NA NArx=2 13 5 422 7.9 365 353 NA> summary(sfit)Call: survfit.coxph(object = fit, newdata = list(age = c(30, 70), ecog.ps = c(2, 3)))rx=1time n.risk n.event survival1 survival259 13 1 0.999 0.87905115 12 1 0.999 0.74575156 11 1 0.998 0.57399268 10 1 0.996 0.41765329 9 1 0.992 0.16677431 8 1 0.988 0.06492638 5 1 0.973 0.00161rx=2time n.risk n.event survival1 survival2353 13 1 0.999 0.7093365 12 1 0.997 0.4739464 9 1 0.994 0.2494475 8 1 0.991 0.1207563 7 1 0.987 0.0489> detach(w=2)> > > # Test the robust=T option of coxph> fit <- coxph(Surv(futime, fustat) ~ age + ecog.ps + rx, ovarian, robust=T)> rr <- resid(fit, type="dfbeta")> all.equal(as.vector(t(rr) %*% rr), as.vector(fit$var))[1] T> temp _ scan("data.jasa", what=list(id=0, b.mo=0, b.d=0, b.y=0,+ a.mo=0, a.d=0, a.y=0,+ t.mo=0, t.d=0, t.y=0,+ f.mo=0, f.d=0, f.y=0,+ fu.stat=0, surg=0, mismatch=0,+ hla.a2=0, mscore=0, reject=0))> > temp3 _ mdy.date(temp$b.mo, temp$b.d, temp$b.y)> temp4 _ mdy.date(temp$a.mo, temp$a.d, temp$a.y)> temp5 _ mdy.date(temp$t.mo, temp$t.d, temp$t.y)> temp6 _ mdy.date(temp$f.mo, temp$f.d, temp$f.y)> > # Make sure that a particular idiocy is turned off: turning logicals into> # factors!> as.data.frame.logical <- as.data.frame.vectorWarning messages:assigning "as.data.frame.logical" masks an object of the same name ondatabase 5> jasa <- data.frame( birth.dt=temp3, accept.dt=temp4, tx.date=temp5,+ fu.date=temp6, fustat=temp$fu.stat,+ surgery = temp$surg, age= temp4-temp3,+ futime = 1+temp6 - temp4, wait.time= 1+temp5 - temp4,+ transplant = c(!is.na(temp5)),+ mismatch=temp$mismatch, hla.a2=temp$hla.a2,+ mscore = temp$mscore, reject=temp$reject)> > row.names(jasa) <- temp$id> > # The "1+" above causes us to match the analysis in Kalbfleisch and Prentice.> # Someone accepted and transplanted on the same day is assumed to have one> # day under the non-transplant risk.>> rm(temp, temp3, temp4, temp5, temp6)> > > > > expect <- survexp(futime ~ ratetable(age=(accept.dt - birth.dt), sex=1,+ year=accept.dt), jasa, cohort=F, ratetable=survexp.uswhite)> > survdiff(Surv(jasa$futime, jasa$fustat) ~ offset(expect))Call:survdiff(formula = Surv(jasa$futime, jasa$fustat) ~ offset(expect))Observed Expected Z p75 0.583 -96.9 0> # Do a Stanford heart transplant data the way that K&P do> #> # Input - a data frame containing the raw data> # Output- a data frame with multiple obs for the transplanted subjects> #> # There are more efficient ways to do this than the "for" loop below, but> # this script is much more readable than any of them.> #> stan1 <- function(jasa) {+ id <- row.names(jasa);+ tx _ 1*(jasa$transplant);+ covar _ cbind(jasa$age/365.25 -48,+ (jasa$accept.dt - mdy.date(10,1,1967))/365.25,+ jasa$surgery);+ n _ length(tx) #number in study;+ ntx _ sum(tx) #number transplanted;+ # Per paragraph 2, p138, the patient who died on the day of transplant is+ # treated differently, for all others deaths are assumed to occur "earlier+ # in the day" than transplants.+ special <- id ==38;+ wait _ jasa$wait.time;+ wait[special] _ wait[special] - .5;++ age <- year <- surgery <- transplant <- id2 <- double(n+ntx);+ start <- stop <- event <- double(n+ntx);+ ii <- 1;+ for (i in 1:n) {+ age[ii] <- covar[i,1];+ year[ii] <- covar[i,2];+ surgery[ii] <- covar[i,3];+ transplant[ii] <- 0;+ id2[ii] <- id[i];++ if (tx[i]) { #transplanted - 2 lines if data+ start[ii] <- 0;+ stop[ii] <- wait[i];+ event[ii] <- 0;++ ii <- ii+1;+ start[ii] <- wait[i];+ stop[ii] <- jasa$futime[i];+ event[ii] <- jasa$fustat[i];+ age[ii] <- covar[i,1];+ year[ii] <- covar[i,2];+ surgery[ii] <- covar[i,3];+ transplant[ii] <- 1;+ id2[ii] <- id[i];+ };+ else { # one line of data+ start[ii] <-0;+ stop[ii] <- jasa$futime[i];+ event[ii]<- jasa$fustat[i];+ };+ ii <- ii+1;+ };++ data.frame(start, stop, event, age, year, surgery,+ transplant=factor(transplant), id=id2);+ }> > jasa1 _ stan1(jasa)> attach(jasa1)> # Now fit the 6 models found in Kalbfleisch and Prentice, p139> options(contrasts=c("contr.treatment", "contr.treatment"))> sfit.1 _ coxph(Surv(start, stop, event)~ (age + surgery)*transplant,+ method="breslow")> sfit.2 _ coxph(Surv(start, stop, event)~ year*transplant,+ method="breslow")> sfit.3 _ coxph(Surv(start, stop, event)~ (age + year)*transplant,+ method="breslow")> sfit.4 _ coxph(Surv(start, stop, event)~ (year +surgery) *transplant,+ method="breslow")> sfit.5 _ coxph(Surv(start, stop, event)~ (age + surgery)*transplant + year ,+ method="breslow")> sfit.6 _ coxph(Surv(start, stop, event)~ age*transplant + surgery + year,+ method="breslow")> > summary(sfit.1)Call:coxph(formula = Surv(start, stop, event) ~ (age + surgery) * transplant, method= "breslow")n= 172coef exp(coef) se(coef) z page 0.0138 1.014 0.0181 0.763 0.45surgery -0.5457 0.579 0.6109 -0.893 0.37transplant 0.1181 1.125 0.3277 0.360 0.72age:transplant 0.0348 1.035 0.0273 1.276 0.20surgery:transplant -0.2916 0.747 0.7582 -0.385 0.70exp(coef) exp(-coef) lower .95 upper .95age 1.014 0.986 0.979 1.05surgery 0.579 1.726 0.175 1.92transplant 1.125 0.889 0.592 2.14age:transplant 1.035 0.966 0.982 1.09surgery:transplant 0.747 1.339 0.169 3.30Rsquare= 0.07 (max possible= 0.969 )Likelihood ratio test= 12.4 on 5 df, p=0.0291Wald test = 11.6 on 5 df, p=0.0402Score (logrank) test = 12 on 5 df, p=0.0345> sfit.2Call:coxph(formula = Surv(start, stop, event) ~ year * transplant, method ="breslow")coef exp(coef) se(coef) z pyear -0.265 0.767 0.105 -2.518 0.012transplant -0.282 0.754 0.514 -0.549 0.580year:transplant 0.136 1.146 0.141 0.967 0.330Likelihood ratio test=8.61 on 3 df, p=0.035 n= 172> summary(sfit.3)Call:coxph(formula = Surv(start, stop, event) ~ (age + year) * transplant, method ="breslow")n= 172coef exp(coef) se(coef) z page 0.0155 1.016 0.0173 0.895 0.3700year -0.2735 0.761 0.1058 -2.585 0.0097transplant -0.5884 0.555 0.5427 -1.084 0.2800age:transplant 0.0339 1.034 0.0279 1.211 0.2300year:transplant 0.2013 1.223 0.1425 1.413 0.1600exp(coef) exp(-coef) lower .95 upper .95age 1.016 0.985 0.982 1.051year 0.761 1.315 0.618 0.936transplant 0.555 1.801 0.192 1.609age:transplant 1.034 0.967 0.979 1.093year:transplant 1.223 0.818 0.925 1.617Rsquare= 0.083 (max possible= 0.969 )Likelihood ratio test= 14.8 on 5 df, p=0.0111Wald test = 13.8 on 5 df, p=0.0172Score (logrank) test = 14 on 5 df, p=0.0153> sfit.4Call:coxph(formula = Surv(start, stop, event) ~ (year + surgery) * transplant,method = "breslow")coef exp(coef) se(coef) z pyear -0.254 0.776 0.108 -2.356 0.018surgery -0.236 0.790 0.628 -0.376 0.710transplant -0.292 0.747 0.506 -0.577 0.560year:transplant 0.164 1.179 0.142 1.162 0.250surgery:transplant -0.550 0.577 0.776 -0.709 0.480Likelihood ratio test=12.3 on 5 df, p=0.0303 n= 172> sfit.5Call:coxph(formula = Surv(start, stop, event) ~ (age + surgery) * transplant + year,method = "breslow")coef exp(coef) se(coef) z page 0.0150 1.015 0.0176 0.851 0.390surgery -0.4192 0.658 0.6157 -0.681 0.500transplant 0.0772 1.080 0.3316 0.233 0.820year -0.1363 0.873 0.0710 -1.921 0.055age:transplant 0.0270 1.027 0.0271 0.995 0.320surgery:transplant -0.2981 0.742 0.7580 -0.393 0.690Likelihood ratio test=16.2 on 6 df, p=0.0127 n= 172> sfit.6Call:coxph(formula = Surv(start, stop, event) ~ age * transplant + surgery + year,method = "breslow")coef exp(coef) se(coef) z page 0.0152 1.015 0.0175 0.870 0.380transplant 0.0475 1.049 0.3222 0.147 0.880surgery -0.6212 0.537 0.3679 -1.689 0.091year -0.1361 0.873 0.0709 -1.919 0.055age:transplant 0.0271 1.027 0.0271 0.998 0.320Likelihood ratio test=16.1 on 5 df, p=0.00667 n= 172> > # Survival curve for the "average" subject> summary(survfit(sfit.1))Call: survfit.coxph(object = sfit.1)time n.risk n.event survival std.err lower 95% CI upper 95% CI1 103 1 0.991 0.00941 0.9723 1.0002 102 3 0.963 0.01879 0.9268 1.0003 99 3 0.935 0.02493 0.8876 0.9855 96 2 0.917 0.02823 0.8629 0.9746 94 2 0.898 0.03113 0.8390 0.9618 92 1 0.889 0.03246 0.8273 0.9559 91 1 0.879 0.03370 0.8158 0.94812 89 1 0.870 0.03485 0.8043 0.94116 88 3 0.842 0.03783 0.7710 0.92017 85 1 0.832 0.03877 0.7598 0.91218 84 1 0.823 0.03959 0.7492 0.90521 83 2 0.805 0.04089 0.7290 0.89028 81 1 0.796 0.04144 0.7190 0.88230 80 1 0.787 0.04196 0.7090 0.87432 78 1 0.778 0.04245 0.6989 0.86635 77 1 0.769 0.04289 0.6890 0.85736 76 1 0.759 0.04330 0.6790 0.84937 75 1 0.750 0.04370 0.6691 0.84139 74 1 0.741 0.04406 0.6593 0.83240 72 2 0.722 0.04478 0.6396 0.81643 70 1 0.713 0.04512 0.6298 0.80745 69 1 0.704 0.04547 0.6199 0.79950 68 1 0.694 0.04580 0.6101 0.79051 67 1 0.685 0.04613 0.6003 0.78253 66 1 0.676 0.04645 0.5905 0.77358 65 1 0.666 0.04678 0.5807 0.76561 64 1 0.657 0.04711 0.5708 0.75666 63 1 0.647 0.04745 0.5609 0.74768 62 2 0.628 0.04812 0.5408 0.73069 60 1 0.619 0.04847 0.5306 0.72172 59 2 0.600 0.04916 0.5106 0.70477 57 1 0.590 0.04952 0.5003 0.69578 56 1 0.580 0.04988 0.4897 0.68680 55 1 0.569 0.05026 0.4790 0.67781 54 1 0.559 0.05062 0.4682 0.66885 53 1 0.549 0.05101 0.4573 0.65890 52 1 0.538 0.05140 0.4464 0.64996 51 1 0.528 0.05178 0.4353 0.639100 50 1 0.517 0.05217 0.4239 0.630102 49 1 0.506 0.05255 0.4125 0.620110 47 1 0.495 0.05295 0.4010 0.610149 45 1 0.483 0.05337 0.3894 0.600153 44 1 0.472 0.05380 0.3778 0.590165 43 1 0.461 0.05423 0.3661 0.581186 41 1 0.450 0.05466 0.3545 0.571188 40 1 0.439 0.05504 0.3430 0.561207 39 1 0.427 0.05541 0.3316 0.551219 38 1 0.416 0.05579 0.3200 0.541263 37 1 0.405 0.05614 0.3083 0.531285 35 2 0.381 0.05678 0.2850 0.511308 33 1 0.370 0.05705 0.2733 0.500334 32 1 0.358 0.05732 0.2616 0.490340 31 1 0.346 0.05755 0.2500 0.480343 29 1 0.334 0.05785 0.2378 0.469584 21 1 0.317 0.05882 0.2202 0.456675 17 1 0.297 0.06009 0.1997 0.442733 16 1 0.276 0.06110 0.1792 0.426852 14 1 0.253 0.06221 0.1562 0.410980 11 1 0.226 0.06332 0.1306 0.391996 10 1 0.201 0.06314 0.1084 0.3721032 9 1 0.177 0.06190 0.0890 0.3511387 6 1 0.150 0.05998 0.0683 0.328> > # Survival curve for a subject of age 50, with prior surgery, tx at 6 months> data <- data.frame(start=c(0,183), stop=c(183,3*365), event=c(1,1),+ age=c(50,50), surgery=c(1,1), transplant=c(0,1))> summary(survfit(sfit.1, data, individual=T))Call: survfit.coxph(object = sfit.1, newdata = data, individual = T)time n.risk n.event survival std.err lower 95% CI upper 95% CI1 103 1 0.98769 0.01780 9.53e-01 12 102 3 0.95150 0.05540 8.49e-01 13 99 3 0.91564 0.09097 7.54e-01 15 96 2 0.89180 0.11384 6.94e-01 16 94 2 0.86811 0.13588 6.39e-01 18 92 1 0.85629 0.14663 6.12e-01 19 91 1 0.84449 0.15722 5.86e-01 112 89 1 0.83264 0.16772 5.61e-01 116 88 3 0.79759 0.19763 4.91e-01 117 85 1 0.78574 0.20739 4.68e-01 118 84 1 0.77425 0.21674 4.47e-01 121 83 2 0.75216 0.23437 4.08e-01 128 81 1 0.74102 0.24304 3.90e-01 130 80 1 0.72986 0.25157 3.71e-01 132 78 1 0.71859 0.26004 3.54e-01 135 77 1 0.70740 0.26836 3.36e-01 136 76 1 0.69623 0.27649 3.20e-01 137 75 1 0.68508 0.28437 3.04e-01 139 74 1 0.67395 0.29214 2.88e-01 140 72 2 0.65191 0.30690 2.59e-01 143 70 1 0.64092 0.31403 2.45e-01 145 69 1 0.62985 0.32100 2.32e-01 150 68 1 0.61889 0.32775 2.19e-01 151 67 1 0.60798 0.33423 2.07e-01 153 66 1 0.59720 0.34046 1.95e-01 158 65 1 0.58633 0.34647 1.84e-01 161 64 1 0.57549 0.35229 1.73e-01 166 63 1 0.56462 0.35789 1.63e-01 168 62 2 0.54280 0.36848 1.43e-01 169 60 1 0.53181 0.37344 1.34e-01 172 59 2 0.51031 0.38245 1.17e-01 177 57 1 0.49939 0.38664 1.10e-01 178 56 1 0.48819 0.39065 1.02e-01 180 55 1 0.47691 0.39443 9.43e-02 181 54 1 0.46556 0.39792 8.72e-02 185 53 1 0.45416 0.40120 8.04e-02 190 52 1 0.44281 0.40416 7.40e-02 196 51 1 0.43131 0.40682 6.79e-02 1100 50 1 0.41964 0.40913 6.21e-02 1102 49 1 0.40796 0.41106 5.66e-02 1110 47 1 0.39623 0.41261 5.15e-02 1149 45 1 0.38451 0.41382 4.66e-02 1153 44 1 0.37287 0.41461 4.22e-02 1165 43 1 0.36123 0.41506 3.80e-02 1186 41 1 0.30934 0.36341 3.09e-02 1188 40 1 0.26422 0.32398 2.39e-02 1207 39 1 0.22454 0.29254 1.75e-02 1219 38 1 0.18973 0.26639 1.21e-02 1263 37 1 0.15917 0.24323 7.96e-03 1285 35 2 0.10980 0.20117 3.03e-03 1308 33 1 0.09033 0.18138 1.76e-03 1334 32 1 0.07368 0.16215 9.86e-04 1340 31 1 0.05969 0.14362 5.34e-04 1343 29 1 0.04755 0.12524 2.72e-04 1584 21 1 0.03414 0.10194 9.82e-05 1675 17 1 0.02272 0.07786 2.75e-05 1733 16 1 0.01448 0.05684 6.61e-06 1852 14 1 0.00828 0.03777 1.08e-06 1980 11 1 0.00409 0.02195 1.11e-07 1996 10 1 0.00194 0.01206 9.82e-09 11032 9 1 0.00087 0.00622 7.14e-10 1> > # These should all give the same answer> j.age <- jasa$age/365.25 -48> fit1 <- coxph(Surv(futime, fustat) ~ j.age, data=jasa)> fit2 <- coxph(Surv(futime, fustat) ~ j.age, jasa, init=fit1$coef, iter=0)> fit3 <- coxph(Surv(start, stop, event) ~ age)> fit4 <- coxph(Surv(start, stop, event) ~ offset((age-fit3$means)*fit1$coef))> s1 <- survfit(fit1, fit3$means)> s2 <- survfit(fit2, fit3$means)> s3 <- survfit(fit3)> s4 <- survfit(fit4)> > all.equal(s1$surv, s2$surv)[1] T> all.equal(s1$surv, s3$surv)[1] T> all.equal(s1$surv, s4$surv)[1] T> > # Still the same answer, fit multiple strata at once> # Strata 1 has independent coefs of strata 2, so putting in> # the other data should not affect it> ll <- length(start)> ss <- rep(0:1, c(ll,ll))> fit <- coxph(Surv(rep(start,2), rep(stop,2), rep(event,2)) ~+ rep(age,2)*strata(ss) + I(rep(age,2)^2*ss) )> fit$coef[1] - fit3$coefrep(age, 2)-1.040834e-17> s4 <- survfit(fit, c(fit3$means, 0,0))> all.equal(s4$surv[1:(s4$strata[1])], s3$surv)[1] T> detach("jasa1")> # A subset of some local data on lung cancer patients. A useful internal test> # because it has multiple strata and lots of missing values.> #>> # inst = enrolling institution> # sex 1=male 2=female> # ph.ecog physician's estimate of the ECOG performace score. 0=fully active,> # 4=bedridden> # ph.karno physician's estimate of the Karnofsky score, a competitor to the> # ECOG ps.> # pat.karno patient's assesment of his/her Karnofsky score> # meal.cal # calories consumed at meals (exclude beverages and snacks)> # wt.loss weight loss in the last 6 months>>> # 12/98: this is now part of the Splus distribution, but they call it> # "lung"> cancer <- lung> #> # Test out all of the routines on a more complex data set> #> temp <- survfit(Surv(time, status) ~ ph.ecog, cancer)> summary(temp, times=c(30*1:11, 365*1:3))Call: survfit(formula = Surv(time, status) ~ ph.ecog, data = cancer)1 observations deleted due to missingph.ecog=0time n.risk n.event survival std.err lower 95% CI upper 95% CI30 60 3 0.952 0.0268 0.9012 1.00060 58 2 0.921 0.0341 0.8562 0.99090 56 2 0.889 0.0396 0.8146 0.970120 56 0 0.889 0.0396 0.8146 0.970150 55 1 0.873 0.0419 0.7946 0.959180 52 2 0.841 0.0461 0.7553 0.936210 48 2 0.808 0.0498 0.7164 0.912240 45 0 0.808 0.0498 0.7164 0.912270 38 2 0.770 0.0543 0.6709 0.884300 33 2 0.727 0.0591 0.6203 0.853330 29 2 0.681 0.0637 0.5670 0.818365 22 6 0.535 0.0728 0.4100 0.699730 5 11 0.193 0.0707 0.0943 0.396ph.ecog=1time n.risk n.event survival std.err lower 95% CI upper 95% CI30 111 2 0.982 0.0124 0.9583 1.00060 110 3 0.956 0.0193 0.9186 0.99490 104 4 0.920 0.0255 0.8718 0.972120 99 5 0.876 0.0310 0.8174 0.939150 93 6 0.823 0.0359 0.7556 0.896180 82 8 0.751 0.0407 0.6756 0.836210 68 9 0.666 0.0450 0.5831 0.760240 57 6 0.604 0.0474 0.5176 0.704270 53 4 0.561 0.0487 0.4729 0.665300 46 3 0.527 0.0495 0.4384 0.633330 40 4 0.480 0.0504 0.3903 0.589365 34 4 0.431 0.0509 0.3417 0.543730 7 21 0.114 0.0388 0.0582 0.222ph.ecog=2time n.risk n.event survival std.err lower 95% CI upper 95% CI30 46 5 0.9000 0.0424 0.82057 0.98760 43 2 0.8600 0.0491 0.76900 0.96290 40 3 0.8000 0.0566 0.69647 0.919120 34 4 0.7174 0.0641 0.60216 0.855150 31 3 0.6541 0.0680 0.53342 0.802180 26 6 0.5275 0.0719 0.40385 0.689210 21 4 0.4431 0.0717 0.32266 0.608240 17 3 0.3766 0.0705 0.26100 0.543270 17 0 0.3766 0.0705 0.26100 0.543300 13 3 0.3102 0.0677 0.20223 0.476330 11 2 0.2624 0.0651 0.16135 0.427365 9 2 0.2147 0.0614 0.12258 0.376730 1 6 0.0371 0.0345 0.00601 0.229ph.ecog=3time n.risk n.event survival std.err lower 95% CI upper 95% CI30 1 0 1 0 1 160 1 0 1 0 1 190 1 0 1 0 1 1> print(temp[2:3])Call: survfit(formula = Surv(time, status) ~ ph.ecog, data = cancer)1 observations deleted due to missingn events mean se(mean) median 0.95LCL 0.95UCLph.ecog=1 113 82 385 27.3 306 268 429ph.ecog=2 50 44 259 31.2 199 156 288> > temp <- survfit(Surv(time, status), cancer, type="fleming",+ conf.int=.9, conf.type="log-log", error="tsiatis")> summary(temp, times=30 *1:5)Call: survfit(formula = Surv(time, status), data = cancer, type = "fleming", conf.int= 0.9, conf.type = "log-log", error = "tsiatis")time n.risk n.event survival std.err lower 90% CI upper 90% CI30 219 10 0.956 0.0135 0.928 0.97460 213 7 0.926 0.0173 0.891 0.95090 201 10 0.882 0.0213 0.842 0.913120 189 10 0.838 0.0244 0.793 0.874150 179 10 0.794 0.0268 0.745 0.834> > temp <- survdiff(Surv(time, status) ~ inst, cancer, rho=.5)> print(temp, digits=6)Call:survdiff(formula = Surv(time, status) ~ inst, data = cancer, rho = 0.5)n=227, 1 observations deleted due to missing.N Observed Expected (O-E)^2/E (O-E)^2/Vinst=1 36 21.190058 17.455181 0.799149708 1.171232977inst=2 5 3.173330 1.964395 0.744007932 0.860140808inst=3 19 10.663476 11.958755 0.140294489 0.200472362inst=4 4 2.245347 3.559344 0.485085848 0.677874608inst=5 9 5.010883 4.500982 0.057765161 0.077128402inst=6 14 8.862602 7.078516 0.449665221 0.582743947inst=7 8 4.445647 4.416133 0.000197254 0.000253632inst=10 4 2.901923 2.223283 0.207150016 0.249077097inst=11 18 7.807867 9.525163 0.309611863 0.422142221inst=12 23 14.009656 12.216768 0.263117640 0.365712493inst=13 20 9.140983 11.863298 0.624699853 0.874238212inst=15 6 3.170744 3.558447 0.042241456 0.057938955inst=16 16 8.870360 9.992612 0.126038005 0.175170113inst=21 13 9.263733 4.460746 5.171484268 6.149354145inst=22 17 8.278566 11.971473 1.139171459 1.645863937inst=26 6 1.627074 3.542694 1.035821659 1.286365543inst=32 7 1.792468 2.679904 0.293869782 0.343966668inst=33 2 0.929177 0.416202 0.632249272 0.676682390Chisq= 15.1 on 17 degrees of freedom, p= 0.590384> > temp <- coxph(Surv(time, status) ~ ph.ecog + ph.karno + pat.karno + wt.loss+ + sex + age + meal.cal + strata(inst), cancer)> summary(temp)Call:coxph(formula = Surv(time, status) ~ ph.ecog + ph.karno + pat.karno + wt.loss +sex + age + meal.cal + strata(inst), data = cancer)n=167 (61 observations deleted due to missing)coef exp(coef) se(coef) z pph.ecog 0.73000 2.075 0.268940 2.714 0.0066ph.karno 0.01305 1.013 0.013736 0.950 0.3400pat.karno -0.01410 0.986 0.009368 -1.505 0.1300wt.loss -0.01488 0.985 0.008481 -1.755 0.0790sex -0.66125 0.516 0.233998 -2.826 0.0047age 0.00509 1.005 0.013729 0.371 0.7100meal.cal -0.00024 1.000 0.000302 -0.794 0.4300exp(coef) exp(-coef) lower .95 upper .95ph.ecog 2.075 0.482 1.225 3.515ph.karno 1.013 0.987 0.986 1.041pat.karno 0.986 1.014 0.968 1.004wt.loss 0.985 1.015 0.969 1.002sex 0.516 1.937 0.326 0.817age 1.005 0.995 0.978 1.033meal.cal 1.000 1.000 0.999 1.000Rsquare= 0.167 (max possible= 0.912 )Likelihood ratio test= 30.6 on 7 df, p=7.38e-05Wald test = 28.1 on 7 df, p=0.000207Score (logrank) test = 30.7 on 7 df, p=7e-05> cox.zph(temp)rho chisq pph.ecog 0.0276 0.1078 0.7427ph.karno 0.1331 2.0018 0.1571pat.karno 0.0250 0.0841 0.7718wt.loss -0.0386 0.2122 0.6451sex 0.0399 0.1800 0.6713age 0.0639 0.5600 0.4543meal.cal 0.1611 3.6945 0.0546GLOBAL NA 9.0115 0.2518> cox.zph(temp, transform="identity")rho chisq pph.ecog 0.0221 0.0688 0.793ph.karno 0.1217 1.6743 0.196pat.karno 0.0302 0.1227 0.726wt.loss -0.0516 0.3790 0.538sex 0.0449 0.2280 0.633age 0.0719 0.7085 0.400meal.cal 0.1808 4.6537 0.031GLOBAL NA 10.0537 0.186> > coxph(Surv(rep(0,length(time)), time, status) ~ ph.ecog + ph.karno + pat.karno+ + wt.loss + sex + age + meal.cal + strata(inst), cancer)Call:coxph(formula = Surv(rep(0, length(time)), time, status) ~ ph.ecog + ph.karno +pat.karno + wt.loss + sex + age + meal.cal + strata(inst), data =cancer)coef exp(coef) se(coef) z pph.ecog 0.73000 2.075 0.268940 2.714 0.0066ph.karno 0.01305 1.013 0.013736 0.950 0.3400pat.karno -0.01410 0.986 0.009368 -1.505 0.1300wt.loss -0.01488 0.985 0.008481 -1.755 0.0790sex -0.66125 0.516 0.233998 -2.826 0.0047age 0.00509 1.005 0.013729 0.371 0.7100meal.cal -0.00024 1.000 0.000302 -0.794 0.4300Likelihood ratio test=30.6 on 7 df, p=7.38e-05 n=167 (61 observations deleted due to missing)> bladder <- read.table("data.bladder4",+ col.names=c("id", "rx", "number", "size", "start", "stop",+ "event", "enum"))> > bladder2 <- bladder[bladder$start< bladder$stop, ]> bladder$start <- NULL> bladder <- bladder[bladder$enum<5,]> #> # Fit the models found in Wei et. al.> #> options(contrasts="contr.treatment")> wfit <- coxph(Surv(stop, event) ~ (rx + size + number)* strata(enum) ++ cluster(id), bladder, method="breslow")> wfitCall:coxph(formula = Surv(stop, event) ~ (rx + size + number) * strata(enum) +cluster(id), data = bladder, method = "breslow")coef exp(coef) se(coef) robust se z prx -0.5176 0.596 0.3158 0.3075 -1.683 0.0920size 0.0679 1.070 0.1012 0.0853 0.796 0.4300number 0.2360 1.266 0.0761 0.0721 3.275 0.0011rxstrata(enum)enum=2 -0.1019 0.903 0.5043 0.3265 -0.312 0.7600rxstrata(enum)enum=3 -0.1823 0.833 0.5579 0.3916 -0.465 0.6400rxstrata(enum)enum=4 -0.1328 0.876 0.6581 0.4968 -0.267 0.7900sizestrata(enum)enum=2 -0.1440 0.866 0.1680 0.1119 -1.287 0.2000sizestrata(enum)enum=3 -0.2792 0.756 0.2086 0.1511 -1.847 0.0650sizestrata(enum)enum=4 -0.2708 0.763 0.2514 0.1856 -1.459 0.1400numberstrata(enum)enum=2 -0.0985 0.906 0.1193 0.1144 -0.861 0.3900numberstrata(enum)enum=3 -0.0662 0.936 0.1298 0.1167 -0.567 0.5700numberstrata(enum)enum=4 0.0930 1.098 0.1465 0.1175 0.792 0.4300Likelihood ratio test=29.4 on 12 df, p=0.00344 n= 340> > # Check the rx coefs versus Wei, et al, JASA 1989> rx <- c(1,4,5,6) # the treatment coefs above> cmat <- diag(4);> cmat[1,] <- 1;> #contrast matrix> wfit$coef[rx] %*% cmat # the coefs in their paper (table 5)[,1] [,2] [,3] [,4][1,] -0.5175702 -0.6194396 -0.6998691 -0.6504161> t(cmat) %*% wfit$var[rx,rx] %*% cmat # var matrix (eqn 3.2)[,1] [,2] [,3] [,4][1,] 0.09455665 0.06017601 0.05677152 0.04376149[2,] 0.06017601 0.13242841 0.13011523 0.11603238[3,] 0.05677152 0.13011523 0.17235798 0.15906935[4,] 0.04376149 0.11603238 0.15906935 0.23978095> > # Anderson-Gill fit> fita <- coxph(Surv(start, stop, event) ~ rx + size + number + cluster(id),+ bladder2, method="breslow")> fitaCall:coxph(formula = Surv(start, stop, event) ~ rx + size + number + cluster(id),data = bladder2, method = "breslow")coef exp(coef) se(coef) robust se z prx -0.4071 0.666 0.2001 0.2418 -1.683 0.0920size -0.0401 0.961 0.0703 0.0722 -0.555 0.5800number 0.1607 1.174 0.0480 0.0569 2.824 0.0047Likelihood ratio test=14.1 on 3 df, p=0.00284 n= 190> > # Prentice fits. Their model 1 a and b are the same> fit1p <- coxph(Surv(stop, event) ~ rx + size + number, bladder2,+ subset=(enum==1), method="breslow")> fit2pa <- coxph(Surv(stop, event) ~ rx + size + number, bladder2,+ subset=(enum==2), method="breslow")> fit2pb <- coxph(Surv(stop-start, event) ~ rx + size + number, bladder2,+ subset=(enum==2), method="breslow")> fit3pa <- coxph(Surv(stop, event) ~ rx + size + number, bladder2,+ subset=(enum==3), method="breslow")> #and etc.> fit1pCall:coxph(formula = Surv(stop, event) ~ rx + size + number, data = bladder2, subset= (enum == 1), method = "breslow")coef exp(coef) se(coef) z prx -0.5176 0.596 0.3158 -1.639 0.1000size 0.0679 1.070 0.1012 0.671 0.5000number 0.2360 1.266 0.0761 3.103 0.0019Likelihood ratio test=9.66 on 3 df, p=0.0216 n= 85> fit2paCall:coxph(formula = Surv(stop, event) ~ rx + size + number, data = bladder2, subset= (enum == 2), method = "breslow")coef exp(coef) se(coef) z prx -0.42419 0.654 0.4022 -1.0547 0.29size -0.12496 0.883 0.1171 -1.0674 0.29number 0.00199 1.002 0.0938 0.0212 0.98Likelihood ratio test=2.02 on 3 df, p=0.569 n= 46> fit2pbCall:coxph(formula = Surv(stop - start, event) ~ rx + size + number, data = bladder2,subset = (enum == 2), method = "breslow")coef exp(coef) se(coef) z prx -0.25910 0.772 0.4051 -0.6396 0.52size -0.11621 0.890 0.1192 -0.9748 0.33number -0.00571 0.994 0.0967 -0.0591 0.95Likelihood ratio test=1.27 on 3 df, p=0.735 n= 46> fit3paCall:coxph(formula = Surv(stop, event) ~ rx + size + number, data = bladder2, subset= (enum == 3), method = "breslow")coef exp(coef) se(coef) z prx -0.8985 0.407 0.554 -1.623 0.10size 0.0850 1.089 0.209 0.407 0.68number -0.0172 0.983 0.128 -0.134 0.89Likelihood ratio test=4.16 on 3 df, p=0.245 n= 27> #> # A test of NULL models> #> fit1 <- coxph(Surv(stop, event) ~ rx + strata(number), bladder, iter=0)> fit2 <- coxph(Surv(stop, event) ~ strata(number), bladder)> > all.equal(fit1$loglik[2], fit2$loglik)[1] T> all.equal(fit1$resid, fit2$resid)[1] T> > > fit1 <- coxph(Surv(start, stop, event) ~ rx + strata(number), bladder2, iter=0)> fit2 <- coxph(Surv(start, stop, event) ~ strata(number), bladder2)> > all.equal(fit1$loglik[2], fit2$loglik)[1] T> all.equal(fit1$resid, fit2$resid)[1] T> # Tests of expected survival>> #> # Simple case 1: a single male subject, born 6/6/36 and entered on study 6/6/55.> #> # Compute the 1, 5, 10 and 12 year expected survival>> temp1 <- mdy.date(6,6,36)> temp2 <- mdy.date(6,6,55)> exp1 <- survexp(~ratetable(year=temp2, age=(temp2-temp1), sex=1),+ ratetable=survexp.uswhite,times=c(366, 1827, 3653, 4383))> > # Well, almost easy. The uswhite table assumes that someone age 19 will have> # seen 5 leap years-- but this lad has only seen 4. Thus the routine sees> # him as 1 day shy of his birthday.> # (The age 19 category starts at 6940, but he is 6939 days old at entry).> # Thus his passage through the table is a bit more complicated>> # First epoch: 1 day at age 18, 1954 6/6/55> # 365 days at age 19, 1955 6/7/55 - 6/5/56>> # Second 365 days at age 20, 1956 6/6/56 - 6/5/57> # 365 days at age 21, 1957 6/6/57 - 6/5/58> # 366 days at age 22, 1958 6/6/58 - 6/6/59> # 365 days at age 23, 1959 6/7/59 - 6/5/60>> # Third 365 days at age 24, 1960 6/6/60 - 6/5/61> # 365 days at age 25, 1961 6/6/61 - 6/5/62> # 366 days at age 26, 1962 6/6/62 - 6/6/63> # 365 days at age 27, 1963 6/7/63 - 6/5/64> # 365 days at age 28, 1964 6/6/64 - 6/5/64>> # Fourth 365 days at age 29, 1965 6/6/65 - 6/5/66> # 365 days at age 30, 1966 6/6/66 - 6/5/67>> # remember, a first subscript of "1" is for age 0>> xx <- survexp.uswhite[,1,]> check <- c( (.6*xx[19,1] + .4*xx[19,2]) ++ 365* (.5*xx[20,1] + .5*xx[20,2]) ,++ 365* (.4*xx[21,1] + .6*xx[21,2]) ++ 365* (.3*xx[22,1] + .7*xx[22,2]) ++ 366* (.2*xx[23,1] + .8*xx[23,2]) ++ 365* (.1*xx[24,1] + .9*xx[24,2]) ,++ 365* ( xx[25,2] ) ++ 365* (.9*xx[26,2] + .1*xx[26,3]) ++ 366* (.8*xx[27,2] + .2*xx[27,3]) ++ 365* (.7*xx[28,2] + .3*xx[28,3]) ++ 365* (.6*xx[29,2] + .4*xx[29,3]) ,++ 365* (.5*xx[30,2] + .5*xx[30,3]) ++ 365* (.4*xx[31,2] + .6*xx[31,3]) )> > print(exp1$surv)[1] 0.9984869 0.9917560 0.9840415 0.9808834> print(exp(-cumsum(check)))[1] 0.9984871 0.9917561 0.9840414 0.9808833> > # This does not pass the "all.equal" test. Because of leap year (again),> # the internal S code does not do exactly what is stated above. US> # rate tables are special: the entry for age 20 in 1950 is the probability> # that someone who becomes 20 years old in 1950 will reach his 21st birthday.> # In order to fit this into the general "cutpoints" calculations of> # person-years, dates are adjusted so that everyone appears to have a> # birthday on Jan 1. But because of leap years, a birthday then moves to> # Dec 31 sometimes -- this happens in each of the 366 day intervals above,> # e.g., 1 day at age 22 with 1957 rates, 365 at age 22 with 1958 rates.> # Upshot: they only agree to 6 decimals. Close enough for me!>> rm(temp1, temp2, exp1, check, xx)> # Tests of expected survival>> #> # Simple case 2: make sure that the averages are correct, for Ederer method> #> # Compute the 1, 5, 10 and 12 year expected survival>> temp1 <- mdy.date(1:6,6:11,36:41)> temp2 <- mdy.date(6:1,11:6,55:50)> age <- temp2 - temp1> > exp1 <- survexp(~ratetable(year=temp2, age=(temp2-temp1), sex=1),+ times=c(366, 1827, 3653, 4383))> exp2 <- survexp(~ratetable(year=temp2, age=(temp2-temp1), sex=1) + I(1:6),+ times=c(366, 1827, 3653, 4383))> > print(all.equal(exp1$surv, apply(exp2$surv, 1, mean)))[1] T> > #> # Check that adding more time points doesn't change things> #> exp3 <- survexp(~ratetable(year=temp2, age=(temp2-temp1), sex=1),+ times=sort(c(366, 1827, 3653, 4383, 30*(1:100))))> > exp1$surv[1] 0.9989593 0.9939559 0.9859462 0.9823106> exp3$surv[match(exp1$time, exp3$time, nomatch=0)][1] 0.9989593 0.9939559 0.9859462 0.9823106> > > #> # Now test Hakulinen's method, assuming an analysis date of 3/1/57> #> futime <- mdy.date(3,1,57) - temp2> xtime <- sort(c(futime, 30, 60, 185, 365))> > exp1 <- survexp(futime ~ ratetable(year=temp2, age=(temp2-temp1), sex=1),+ times=xtime, conditional=F)> exp2 <- survexp(~ratetable(year=temp2, age=(temp2-temp1), sex=1) + I(1:6),+ times=futime)> > wt <- rep(1,6)> con <- double(6)> for (i in 1:6) {+ con[i] <- sum(exp2$surv[i,i:6])/sum(wt[i:6]);+ wt <- exp2$surv[i,];+ }> > exp1$surv[match(futime, xtime)][1] 0.9981627 0.9970156 0.9959044 0.9948167 0.9937561 0.9927036> cumprod(con) #should be equal[1] 0.9981627 0.9970156 0.9959044 0.9948167 0.9937561 0.9927036> > > #> # Now for the conditional method> #> exp1 <- survexp(futime ~ ratetable(year=temp2, age=(temp2-temp1), sex=1),+ times=xtime, conditional=T)> > cond <- exp2$surv> for (i in 6:2) cond[i,] _ (cond[i,]/cond[i-1,]) #conditional survival> for (i in 1:6) con[i] _ exp(mean(log(cond[i, i:6])))> > all.equal(exp1$surv[match(futime, xtime)], cumprod(con))[1] T> cumprod(con)[1] 0.9981625 0.9970151 0.9959036 0.9948157 0.9937550 0.9927025> > rm(con, cond, exp1, exp2, wt, temp1, temp2, age)> rm(exp3, futime, xtime)> #> # Test out expected survival, when the parent pop is another Cox model> #> fit <- coxph(Surv(time, status) ~x, test1, method="breslow")> > dummy <- data.frame(time=c(.5, 1, 1.5, 2, 2.5, 3, 3.5, 4, 4.5),+ status=c(1,0,1,0,1,0,1,1,1), x=(-4:4)/2)> > efit <- survexp(time ~ ratetable(x=x), dummy, ratetable=fit, cohort=F)> > #> # Now, compare to the true answer, which is known to us> #> ss <- exp(fit$coef)> haz <- c( 1/(3*ss+3), 2/(ss+3), 1) #truth at time 0,1,2,4+> chaz <- cumsum(c(0,haz))> chaz2 <- chaz[c(1,2,2,3,3,3,3,4,4)]> > risk <- exp(fit$coef*dummy$x)> efit2 <- exp(-risk*chaz2)> > all.equal(as.vector(efit), as.vector(efit2)) #ignore mismatched name attrib[1] T> > #> # Now test the direct-adjusted curve (Ederer)> #> efit <- survexp( ~ ratetable(x=x), dummy, ratetable=fit, se=F)> direct <- survfit(fit, newdata=dummy)$surv> > chaz <- chaz[-1] #drop time 0> d2 <- exp(outer(-chaz, risk))> all.equal(as.vector(direct), as.vector(d2)) #this tests survfit[1] T> > all.equal(as.vector(efit$surv), as.vector(apply(direct,1,mean))) #direct[1] T> > > #> # Now test out the Hakulinen method (Bonsel's method)> # By construction, we have a large correlation between x and censoring> #> # In theory, hak1 and hak2 would be the same. In practice, like a KM and> # F-H, they differ when n is small.> #> efit <- survexp( time ~ ratetable(x=x), dummy, ratetable=fit, se=F)> > surv <- wt <- rep(1,9)> tt <- c(1,2,4)> hak1 <- hak2 <- NULL> for (i in 1:3) {+ wt[dummy$time < tt[i]] <- 0;+ hak1 <- c(hak1, exp(-sum(haz[i]*risk*surv*wt)/sum(surv*wt)));+ hak2 <- c(hak2, sum(exp(-haz[i]*risk)*surv*wt)/sum(surv*wt));+ surv <- surv * exp(-haz[i]*risk);+ }> > all.equal(as.vector(efit$surv), as.vector(cumprod(hak2)))[1] T> > #> # Now do the conditional estimate> #> efit <- survexp( time ~ ratetable(x=x), dummy, ratetable=fit, se=F,+ conditional=T)> wt <- rep(1,9)> cond <- NULL> for (i in 1:3) {+ wt[dummy$time < tt[i]] <- 0;+ cond <- c(cond, exp(-sum(haz[i]*risk*wt)/sum(wt)));+ }> > all.equal(as.vector(efit$surv), as.vector(cumprod(cond)))[1] T> > rm(wt, cond, efit, tt, surv, hak1, hak2)> rm(fit, dummy, ss, efit2, chaz, chaz2, risk)> rm(d2, direct)> #> # Run a test that can be verified using SAS's LIFEREG> #> fit1w <- survreg(Surv(time, status) ~x, test1, dist="weibull")> fit1wCall:survreg(formula = Surv(time, status) ~ x, data = test1, dist = "weibull")Coefficients:(Intercept) x1.293465 -0.7456443Scale= 0.255742Loglik(model)= -4.9 Loglik(intercept only)= -7.1Chisq= 4.35 on 1 degrees of freedom, p= 0.037n=6 (1 observations deleted due to missing)> summary(fit1w)Call:survreg(formula = Surv(time, status) ~ x, data = test1, dist = "weibull")Value Std. Error z p(Intercept) 1.293 0.181 7.15 8.57e-13x -0.746 0.257 -2.90 3.75e-03Log(scale) -1.364 0.392 -3.48 5.04e-04Scale= 0.256Weibull distributionLoglik(model)= -4.9 Loglik(intercept only)= -7.1Chisq= 4.35 on 1 degrees of freedom, p= 0.037Number of Newton-Raphson Iterations: 5n=6 (1 observations deleted due to missing)Correlation of Coefficients:(Intercept) xx -0.705Log(scale) 0.016 -0.109> > fit1e <- survreg(Surv(time, status) ~x, test1, dist="exp")> fit1eCall:survreg(formula = Surv(time, status) ~ x, data = test1, dist = "exp")Coefficients:(Intercept) x1.504077 -0.8109302Scale fixed at 1Loglik(model)= -8.4 Loglik(intercept only)= -8.7Chisq= 0.64 on 1 degrees of freedom, p= 0.42n=6 (1 observations deleted due to missing)> summary(fit1e)Call:survreg(formula = Surv(time, status) ~ x, data = test1, dist = "exp")Value Std. Error z p(Intercept) 1.504 0.707 2.127 0.0334x -0.811 1.000 -0.811 0.4174Scale fixed at 1Exponential distributionLoglik(model)= -8.4 Loglik(intercept only)= -8.7Chisq= 0.64 on 1 degrees of freedom, p= 0.42Number of Newton-Raphson Iterations: 3n=6 (1 observations deleted due to missing)Correlation of Coefficients:(Intercept)x -0.707> > fit1l <- survreg(Surv(time, status) ~x, test1, dist="loglogistic")> fit1lCall:survreg(formula = Surv(time, status) ~ x, data = test1, dist = "loglogistic")Coefficients:(Intercept) x1.190922 -0.791701Scale= 0.2159941Loglik(model)= -5.3 Loglik(intercept only)= -7.1Chisq= 3.62 on 1 degrees of freedom, p= 0.057n=6 (1 observations deleted due to missing)> summary(fit1l)Call:survreg(formula = Surv(time, status) ~ x, data = test1, dist = "loglogistic")Value Std. Error z p(Intercept) 1.191 0.240 4.97 6.71e-07x -0.792 0.355 -2.23 2.59e-02Log(scale) -1.533 0.378 -4.05 5.13e-05Scale= 0.216Log logistic distributionLoglik(model)= -5.3 Loglik(intercept only)= -7.1Chisq= 3.62 on 1 degrees of freedom, p= 0.057Number of Newton-Raphson Iterations: 4n=6 (1 observations deleted due to missing)Correlation of Coefficients:(Intercept) xx -0.671Log(scale) 0.047 0.040> > fit1g <- survreg(Surv(time, status) ~x, test1, dist="lognormal")> summary(fit1g)Call:survreg(formula = Surv(time, status) ~ x, data = test1, dist = "lognormal")Value Std. Error z p(Intercept) 1.16 0.215 5.39 7.08e-08x -0.77 0.308 -2.50 1.24e-02Log(scale) -1.07 0.344 -3.12 1.80e-03Scale= 0.341Log Normal distributionLoglik(model)= -5 Loglik(intercept only)= -7Chisq= 3.89 on 1 degrees of freedom, p= 0.049Number of Newton-Raphson Iterations: 5n=6 (1 observations deleted due to missing)Correlation of Coefficients:(Intercept) xx -0.681Log(scale) 0.167 -0.016> #> # Do a test with the ovarian data> #> fitfw <- survreg.old(Surv(futime, fustat) ~ age + ecog.ps, ovarian,+ link="log", dist="extreme")> fitfwCall:survreg(formula = formula, data = data, dist = dist, scale = scale)Coefficients:(Intercept) age ecog.ps12.28497 -0.09702669 0.09977342Scale= 0.6032744Loglik(model)= -90 Loglik(intercept only)= -98Chisq= 15.98 on 2 degrees of freedom, p= 0.00034n= 26> > fitfl <- survreg.old(Surv(futime, fustat) ~ age + ecog.ps, ovarian,+ link="log", dist="logistic")> fitflCall:survreg(formula = formula, data = data, dist = dist, scale = scale)Coefficients:(Intercept) age ecog.ps11.50853 -0.08876814 0.09033338Scale= 0.4464063Loglik(model)= -89.5 Loglik(intercept only)= -97.4Chisq= 15.67 on 2 degrees of freedom, p= 4e-04n= 26> > flem2 <- scan("fleming.data2", what=list(ltime=0, rtime=0))> flsurv<- Surv(flem2$ltime, flem2$rtime, type="interval2")> > fitfw2 <- survreg(flsurv ~ age + ecog.ps, ovarian, dist="weibull")> summary(fitfw2)Call:survreg(formula = flsurv ~ age + ecog.ps, data = ovarian, dist = "weibull")Value Std. Error z p(Intercept) 12.3870 1.5981 7.751 9.13e-15age -0.0987 0.0253 -3.898 9.70e-05ecog.ps 0.1013 0.3773 0.269 7.88e-01Log(scale) -0.4788 0.2582 -1.854 6.37e-02Scale= 0.62Weibull distributionLoglik(model)= -50.2 Loglik(intercept only)= -58Chisq= 15.62 on 2 degrees of freedom, p= 0.00041Number of Newton-Raphson Iterations: 5n= 26Correlation of Coefficients:(Intercept) age ecog.psage -0.920ecog.ps -0.213 -0.170Log(scale) 0.444 -0.382 -0.114> > fitfl2 <- survreg(flsurv ~ age + ecog.ps, ovarian, dist="loglogistic")> summary(fitfl2)Call:survreg(formula = flsurv ~ age + ecog.ps, data = ovarian, dist = "loglogistic")Value Std. Error z p(Intercept) 11.5250 1.5220 7.572 3.67e-14age -0.0889 0.0239 -3.721 1.98e-04ecog.ps 0.0858 0.3626 0.237 8.13e-01Log(scale) -0.8067 0.2706 -2.981 2.87e-03Scale= 0.446Log logistic distributionLoglik(model)= -49.8 Loglik(intercept only)= -57.6Chisq= 15.46 on 2 degrees of freedom, p= 0.00044Number of Newton-Raphson Iterations: 4n= 26Correlation of Coefficients:(Intercept) age ecog.psage -0.920ecog.ps -0.293 -0.083Log(scale) 0.377 -0.368 0.018> > fitfg2 <- survreg(flsurv ~ age + ecog.ps, ovarian, dist="lognormal")> summary(fitfg2)Call:survreg(formula = flsurv ~ age + ecog.ps, data = ovarian, dist = "lognormal")Value Std. Error z p(Intercept) 11.1575 1.4301 7.802 6.09e-15age -0.0856 0.0237 -3.614 3.02e-04ecog.ps 0.2081 0.3815 0.545 5.85e-01Log(scale) -0.2334 0.2507 -0.931 3.52e-01Scale= 0.792Log Normal distributionLoglik(model)= -50 Loglik(intercept only)= -57.5Chisq= 15.03 on 2 degrees of freedom, p= 0.00055Number of Newton-Raphson Iterations: 4n= 26Correlation of Coefficients:(Intercept) age ecog.psage -0.906ecog.ps -0.197 -0.213Log(scale) 0.483 -0.440 0.008> #> # A simple test of the singular=ok option> #> test1 <- data.frame(time= c(4, 3,1,1,2,2,3),+ status=c(1,NA,1,0,1,1,0),+ x= c(0, 2,1,1,1,0,0))> > temp <- rep(0:3, rep(7,4))> > stest <- data.frame(start = 10*temp,+ stop = 10*temp + test1$time,+ status = rep(test1$status,4),+ x = c(test1$x+ 1:7, rep(test1$x,3)),+ epoch = rep(1:4, rep(7,4)))> > fit1 <- coxph(Surv(start, stop, status) ~ x * factor(epoch), stest)Warning messages:X matrix deemed to be singular; variable 2 3 4 in: coxph(Surv(start, stop,status) ~ x * factor(epoch), stest)> fit1$coef # elements 2:4 should be NAx factor(epoch)2 factor(epoch)3 factor(epoch)4 xfactor(epoch)20.1041579 NA NA NA 1.572694xfactor(epoch)3 xfactor(epoch)41.572694 1.572694> #> # Test some more features of surv.diff> #> # First, what happens when one group is a dummy> #>>> #> # The AML data, with a third group of early censorings "tacked on"> #> aml3 <- list(time= c( 9, 13, 13, 18, 23, 28, 31, 34, 45, 48, 161,+ 5, 5, 8, 8, 12, 16, 23, 27, 30, 33, 43, 45,+ 1, 2, 2, 3, 3, 3, 4),+ status= c( 1,1,0,1,1,0,1,1,0,1,0, 1,1,1,1,1,0,1,1,1,1,1,1,+ 0,0,0,0,0,0,0),+ x = as.factor(c(rep("Maintained", 11),+ rep("Nonmaintained", 12), rep("Dummy",7) )))> > aml3 <- data.frame(aml3)> > # These should give the same result (chisq, df), but the second has an> # extra group> survdiff(Surv(time, status) ~x, aml)Call:survdiff(formula = Surv(time, status) ~ x, data = aml)N Observed Expected (O-E)^2/E (O-E)^2/Vx=Maintained 11 7 10.69 1.27 3.4x=Nonmaintained 12 11 7.31 1.86 3.4Chisq= 3.4 on 1 degrees of freedom, p= 0.0653> survdiff(Surv(time, status) ~x, aml3)Call:survdiff(formula = Surv(time, status) ~ x, data = aml3)N Observed Expected (O-E)^2/E (O-E)^2/Vx=Dummy 7 0 0.00 NA NAx=Maintained 11 7 10.69 1.27 3.4x=Nonmaintained 12 11 7.31 1.86 3.4Chisq= 3.4 on 1 degrees of freedom, p= 0.0653> > > #> # Now a test of the stratified log-rank> # There are no tied times within institution, so the coxph program> # can be used to give a complete test> #> fit <- survdiff(Surv(time, status) ~ pat.karno + strata(inst), cancer)> > options(contrasts="contr.treatment")> cfit <- coxph(Surv(time, status) ~ factor(pat.karno) + strata(inst),+ cancer, iter=0)> > tdata <- na.omit(cancer[,c("time", "status", "pat.karno", "inst")])> > temp1 <- tapply(tdata$status-1, list(tdata$pat.karno, tdata$inst), sum)> temp1 <- ifelse(is.na(temp1), 0, temp1)> temp2 <- tapply(cfit$resid, list(tdata$pat.karno, tdata$inst), sum)> temp2 <- ifelse(is.na(temp2), 0, temp2)> > temp2 <- temp1 - temp2> > #Now temp1=observed, temp2=expected> all.equal(c(temp1), c(fit$obs))[1] T> all.equal(c(temp2), c(fit$exp))[1] T> > all.equal(fit$var[-1,-1], solve(cfit$var))[1] T> > rm(tdata, temp1, temp2)> #> # Simple case: a single male subject, born 6/6/36 and entered on study 6/6/55.> #>> temp1 <- mdy.date(6,6,36)> temp2 <- mdy.date(6,6,55)# Now compare the results from person-years> #> temp.age <- tcut(temp2-temp1, floor(c(-1, (18:31 * 365.24))),+ labels=c("0-18", paste(18:30, 19:31, sep="-")))> temp.yr <- tcut(temp2, mdy.date(1,1,1954:1965), labels=1954:1964)> temp.time <- 3700 #total days of fu> py1 <- pyears(temp.time ~ temp.age + temp.yr, scale=1) #output in days> > # The subject should appear in 20 cells> # 6/6/55 - 12/31/55, 209 days, age 19-20, 1955> # 1/1/56 - 6/ 4/56, 156 days, age 19-20, 1956> # 6/5/56 - 12/31/56, 210 days, age 20-21, 1956 (a leap year, and his> # birthday computes one day earlier)> # 1/1/57 - 6/ 5/57, 156 days, age 20-21, 1957> # 6/6/57 - 12/31/57, 209 days, age 21-22, 1957> # and etc> # with 203 days "off table", ie, beyond the last cell of the table> #> # It is a nuisance, but tcut follows 'cut' in that we give the ENDS of> # the intervals, whereas the survival tables use the starts of intervals.> # Thus this breakdown does not match that in doexpect.s> #> xx <- matrix(0, nrow=14, ncol=11)> xx[cbind(3:11, 3:11)] <- 156> xx[cbind(3:12, 2:11)] <- c(209, 210, rep(c(209, 209, 209, 210),2))> dimnames(xx) <- list(c("0-18", paste(18:30, 19:31, sep="-")), 1954:1964)> all.equal(xx, py1$pyears)[1] T> all.equal(203, py1$offtable)[1] T> all.equal(1*(xx>0), py1$n)[1] T> > #> # Now with expecteds> #> py2 <- pyears(temp.time ~ temp.age + temp.yr+ + ratetable(age=temp2-temp1, year=temp2, sex=1),+ scale=1, ratetable=survexp.uswhite ) #output in days> all.equal(xx, py2$pyears)[1] T> all.equal(203, py2$offtable)[1] T> all.equal(1*(xx>0), py2$n)[1] T> > > py3 <- pyears(temp.time ~ temp.age + temp.yr+ + ratetable(age=temp2-temp1, year=temp2, sex=1),+ scale=1, ratetable=survexp.uswhite , expect="pyears")> all.equal(py2$n, py3$n)[1] T> all.equal(py2$pyear, py3$pyear)[1] T> all.equal(py3$n, 1*(py3$expect>0))[1] T> > # Now, compute the py3 result "by hand". Since there is only one person> # it can be derived from py2.> #> xx1 <- py2$expect[py2$n>0] # the hazard over each interval> cumhaz <- cumsum(c(0, xx1[-length(xx1)])) # the cumulative hazard> xx2 <- py3$expect[py3$n>0] # the expected number of person days> xx3 <- py3$pyears[py3$n>0] # the potential number of person days> > # This is the integral of the curve "exp(-haz *t)" over the interval> integral <- xx3 * exp(-cumhaz)* (1- exp(-xx1))/ xx1> # They might not be exactly equal, since the C code tracks changes in the> # rate tables that occur -within- an interval. So try for 7 digits> all.equal(round(integral,4), round(xx2,4))[1] "Mean relative difference: 4.7713e-07"> > # Cut off the bottom of the table, instead of the side> temp.age <- tcut(temp2-temp1, floor(c(-1, (18:27 * 365.24))),+ labels=c("0-18", paste(18:26, 19:27, sep="-")))> > py4 <- eval(py3$call)> all.equal(py4$pyear, py3$pyear[1:10,])[1] T> all.equal(py4$expect, py3$expect[1:10,])[1] T> > rm(xx1, xx2, xx3, cumhaz, temp1, temp2)> rm(temp.age, temp.yr, temp.time, xx)> #> # Test out subscripting in the case of a coxph survival curve> #> fit <- coxph(Surv(time, status) ~ age + sex + meal.cal + strata(ph.ecog),+ data=cancer)> > surv1 <- survfit(fit)> temp <- surv1[2:3]> > which <- cumsum(surv1$strata)> zed <- (which[1]+1):(which[3])> all.equal(surv1$surv[zed], temp$surv)[1] T> all.equal(surv1$time[zed], temp$time)[1] T> > #> # Now a result with a matrix of survival curves> #> dummy <- data.frame(age=c(30,40,60), sex=c(1,2,2), meal.cal=c(500, 1000, 1500))> surv2 <- survfit(fit, newdata=dummy)> > zed <- 1:which[1]> all.equal(surv2$surv[zed,1], surv2[1,1]$surv)[1] T> all.equal(surv2$surv[zed,2], surv2[1,2]$surv)[1] T> all.equal(surv2$surv[zed,3], surv2[1,3]$surv)[1] T> all.equal(surv2$surv[zed, ], surv2[1,1:3]$surv)[1] T> all.equal(surv2$surv[zed], (surv2[1]$surv)[,1])[1] T> all.equal(surv2$surv[zed, ], surv2[1, ]$surv)[1] T> q()