Parametric models with {icpack}

Hein Putter & Paul Eilers

01 juli, 2024

Introduction

The purpose of this vignette is to show how {icpack} can be used to fit certain parametric models, namely with constant hazards, Gompertz models and Weibull models. We will use the ovarian cancer data, because with right-censored data we can easily check the results of {icpack} with other methods. Those other methods do not extend to interval-censored data, while {icpack} does.

For information on the ovarian cancer data we refer to the vignette “Smooth modeling of right-censored survival data with {icpack}”, or the help file. First the package needs to be loaded.

library(icpack)
library(ggplot2)

Constant hazards

Using icfit it is possible to fit proportional hazards models to right-censored and interval-censored data with constant baseline hazards. It is of interest to start with a model without covariates and right-censored data, since then we know what the maximum likelihood of the underlying constant hazard is, and we also have a straightforward expression for its confidence interval. Here is how the model is fit using icfit.

ic_fit_cst <- icfit(Surv(time, d) ~ 1, data = Ova,
                    bdeg=1, pord=1, lambda=1e+06, update_lambda=FALSE)
ic_fit_cst
## Object of class 'icfit'
## Call:
## icfit(formula = Surv(time, d) ~ 1, data = Ova, lambda = 1e+06, 
##     bdeg = 1, pord = 1, update_lambda = FALSE)
## 
## Bins summary: tmin = 0, tmax = 2756, number of bins = 100, bin width = 27.56
## P-splines summary: number of segments = 20, degree = 1, penalty order = 1
## 
## Parameter estimates:
## 
## Effective baseline dimension: 0.04605, log-likelihood: -1299, AIC: 2598
## Smoothness parameter lambda: 1e+06
## Number of iterations: 4
## n = 358, number of events = 266

We use bdeg=1 to use B-splines of degree 1 (constants), pord=1 to penalize first order differences of coefficients of the B-splines, lambda=1e+06 to use an extremely high penalty, and update_lambda=FALSE to make sure that lambda is not updated in the process. Note that the effective dimension of the log-hazard is effectively 1, so the log-hazard and hazard are constant. A plot of the hazard shows a straight line.

plot(ic_fit_cst)

plot of chunk plot_icfit_cstplot of chunk plot_icfit_cst

The value fo the hazard and its 95% confidence interval can be seen by printing the result of predict.

pic_fit_cst <- predict(ic_fit_cst)[, c("time", "haz", "sehaz", "lowerhaz", "upperhaz")]
head(pic_fit_cst)
##       time          haz        sehaz     lowerhaz     upperhaz
## 1  0.00000 0.0007470698 4.582517e-05 0.0006624431 0.0008425075
## 2  5.51258 0.0007470698 4.582473e-05 0.0006624439 0.0008425065
## 3 11.02516 0.0007470699 4.582430e-05 0.0006624447 0.0008425057
## 4 16.53774 0.0007470700 4.582390e-05 0.0006624455 0.0008425049
## 5 22.05032 0.0007470700 4.582352e-05 0.0006624462 0.0008425041
## 6 27.56290 0.0007470701 4.582316e-05 0.0006624469 0.0008425033
tail(pic_fit_cst)
##         time          haz        sehaz     lowerhaz     upperhaz
## 496 2728.727 0.0007468735 4.586438e-05 0.0006621800 0.0008423994
## 497 2734.240 0.0007468735 4.586478e-05 0.0006621793 0.0008424003
## 498 2739.752 0.0007468735 4.586520e-05 0.0006621785 0.0008424012
## 499 2745.265 0.0007468735 4.586564e-05 0.0006621778 0.0008424022
## 500 2750.777 0.0007468735 4.586609e-05 0.0006621770 0.0008424032
## 501 2756.290 0.0007468735 4.586657e-05 0.0006621761 0.0008424042

We see that the hazard is estimated as \(\hat{\lambda} = 0.000747\), with 95% confidence interval about 0.000662 – 0.000842. The estimate and confidence interval are obtained using the general theory outlined in Putter, Gampe & Eilers (2023). Nevertheless, they coincide closely with standard theory, which says that \(\hat{\lambda}_{\textrm{MLE}} = D/T\), where \(D = \sum_{i=1}^n d_i\) is the total number of events and \(T = \sum_{i=1}^n t_i\) is the total follow-up time in the data. (Here \((t_i, d_i)\) for subject \(i=1,\ldots,n\) are the usual time and status indicators in right-censored time-to-event data.) Moreover, the variance of the log of \(\hat{\lambda}_{\textrm{MLE}}\) is given by \(1/D\), so a 95% confidence interval for \(\lambda\) is given by \(\exp( \log(\hat{\lambda}_{\textrm{MLE}}) \pm 1.96/\sqrt{D} )\). In our data this would yield

D <- sum(Ova$d)
T <- sum(Ova$time)
rate <- D / T
lograte <- log(rate)
selograte <- 1 / sqrt(D)
lowerlograte <- lograte - qnorm(0.975) * selograte
upperlograte <- lograte + qnorm(0.975) * selograte
data.frame(D=D, T=T, rate=rate, lower=exp(lowerlograte), upper=exp(upperlograte))
##     D      T         rate        lower        upper
## 1 266 351135 0.0007575434 0.0006717644 0.0008542757

We see that this is quite close to what icfit has given us.

We can go one step further with icfit and add the covariate effects to the constant baseline hazard.

ic_fit_cst <- icfit(Surv(time, d) ~ Diameter + FIGO + Karnofsky, data = Ova,
                    bdeg=1, pord=1, lambda=1e+06, update_lambda=FALSE)
ic_fit_cst
## Object of class 'icfit'
## Call:
## icfit(formula = Surv(time, d) ~ Diameter + FIGO + Karnofsky, 
##     data = Ova, lambda = 1e+06, bdeg = 1, pord = 1, update_lambda = FALSE)
## 
## Bins summary: tmin = 0, tmax = 2756, number of bins = 100, bin width = 27.56
## P-splines summary: number of segments = 20, degree = 1, penalty order = 1
## 
## Parameter estimates:
##               coef       SE       HR    lower upper   pvalue    
## Diameter   0.21013  0.04931  1.23384  1.12017 1.359 2.03e-05 ***
## FIGO       0.56814  0.13461  1.76498  1.35569 2.298 2.44e-05 ***
## Karnofsky -0.17083  0.05325  0.84296  0.75942 0.936  0.00134 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Effective baseline dimension: 0.04602, log-likelihood: -1264, AIC: 2528
## Smoothness parameter lambda: 1e+06
## Number of iterations: 7
## n = 358, number of events = 266

Also this model can be fitted in an alternative way, namely using direct Poisson regression. To achieve this d is considered the outcome variable and the logarithm of the follow-up time is added as an offset. Here is the code to run it.

Ova$logtime <- log(Ova$time)
pois_fit <- glm(d ~ Diameter + FIGO + Karnofsky + offset(logtime), family=poisson(), data = Ova)
spois_fit <- summary(pois_fit)
spois_fit
## 
## Call:
## glm(formula = d ~ Diameter + FIGO + Karnofsky + offset(logtime), 
##     family = poisson(), data = Ova)
## 
## Coefficients:
##             Estimate Std. Error z value Pr(>|z|)    
## (Intercept) -6.33203    0.53325 -11.874  < 2e-16 ***
## Diameter     0.21167    0.04932   4.291 1.77e-05 ***
## FIGO         0.57594    0.13456   4.280 1.87e-05 ***
## Karnofsky   -0.17387    0.05332  -3.261  0.00111 ** 
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for poisson family taken to be 1)
## 
##     Null deviance: 595.55  on 357  degrees of freedom
## Residual deviance: 523.70  on 354  degrees of freedom
## AIC: 1063.7
## 
## Number of Fisher Scoring iterations: 6

We see that the estimates of the covariate effects are quite similar to the estimates from ic_fit_cst. Moreover the (Intercept) term is the logarithm of the (constant) baseline hazard. If we exponeniate it the corresponding baseline hazard and 95% confidence interval are given by

lograte <- spois_fit$coefficients[1, 1]
selograte <- spois_fit$coefficients[1, 2]
lowerlograte <- lograte - qnorm(0.975) * selograte
upperlograte <- lograte + qnorm(0.975) * selograte
data.frame(rate=exp(lograte), lower=exp(lowerlograte), upper=exp(upperlograte))
##          rate        lower       upper
## 1 0.001778422 0.0006253608 0.005057535

The result from icfit is visualised as

plot(ic_fit_cst, xlab="Days since randomisation", ylab="Hazard",
     title="Constant baseline hazard") + ylim(0, 0.0052)

plot of chunk plot_icfit_cst_covsplot of chunk plot_icfit_cst_covs

Naturally, direct Poisson regression using glm is preferable to icfit in case of constant hazards both in terms of speed and accuracy, but it is comforting to see that icfit does give rapid and reliable results also here. Moreover, Poisson regression cannot be used directly with interval-censored data.

Precision in icfit can be increased (at the cost of extra computation time) by increasing nt. If we remove the covariates again, we set nt to 500, and look at the hazard and its 95% confidence interval, which again can be seen by printing the result of predict, we do see that the results are quite a bit closer to that of the occurrence over exposure theoretical ones.

ic_fit_cst <- icfit(Surv(time, d) ~ 1, data = Ova, nt=500,
                    bdeg=1, pord=1, lambda=1e+06, update_lambda=FALSE)
pic_fit_cst <- predict(ic_fit_cst)[, c("time", "haz", "sehaz", "lowerhaz", "upperhaz")]
head(pic_fit_cst)
##       time          haz        sehaz     lowerhaz     upperhaz
## 1  0.00000 0.0007554777 4.634085e-05 0.0006698987 0.0008519894
## 2  5.51258 0.0007554778 4.634041e-05 0.0006698995 0.0008519885
## 3 11.02516 0.0007554779 4.633998e-05 0.0006699003 0.0008519876
## 4 16.53774 0.0007554779 4.633958e-05 0.0006699011 0.0008519868
## 5 22.05032 0.0007554780 4.633919e-05 0.0006699018 0.0008519860
## 6 27.56290 0.0007554780 4.633882e-05 0.0006699025 0.0008519852
tail(pic_fit_cst)
##         time          haz        sehaz     lowerhaz     upperhaz
## 496 2728.727 0.0007552829 4.638097e-05 0.0006696355 0.0008518848
## 497 2734.240 0.0007552829 4.638137e-05 0.0006696348 0.0008518857
## 498 2739.752 0.0007552829 4.638179e-05 0.0006696340 0.0008518866
## 499 2745.265 0.0007552829 4.638223e-05 0.0006696333 0.0008518876
## 500 2750.777 0.0007552829 4.638270e-05 0.0006696325 0.0008518886
## 501 2756.290 0.0007552829 4.638318e-05 0.0006696316 0.0008518897

Note that only with right-censored data we can use the occurrence over exposure theory and Poisson GLM. Importantly, {icpack} can also be used to fit models (with and without covariates) with constant (baseline) hazards for interval-censored data, while there is no standard alternative for this situation.

“Gompertz” models

When we set the degree of the B-splines to 1 and the order of the penalty to 2, and take a very large value of lambda, we get a model for which the log-hazard is a straight line. The corresponding hazard is of the form \(h(t) = a e^{bt}\). This corresponds to the hazard of a Gompertz model, provided \(b>0\). Gompertz rates are often used to represent rates of mortality; when \(b<0\) then this implies that very large, or indeed infinite values are not uncommon, since then \(\int_0^\infty h(t) dt\) is finite. For modeling human mortality this obviously is a problem, but for many other practical purposes we may ignore this issue. But implicitly allowing for negative \(b\) by icfit is the reason for referring to “Gompertz” rather than Gompertz models.

Here is the fit of the “Gompertz” model using icfit.

ic_fit_gompertz <- icfit(Surv(time, d) ~ 1, data = Ova,
                         bdeg=1, pord=2, lambda=1e+06, update_lambda=FALSE)
ic_fit_gompertz
## Object of class 'icfit'
## Call:
## icfit(formula = Surv(time, d) ~ 1, data = Ova, lambda = 1e+06, 
##     bdeg = 1, pord = 2, update_lambda = FALSE)
## 
## Bins summary: tmin = 0, tmax = 2756, number of bins = 100, bin width = 27.56
## P-splines summary: number of segments = 20, degree = 1, penalty order = 2
## 
## Parameter estimates:
## 
## Effective baseline dimension: 1.048, log-likelihood: -1291, AIC: 2585
## Smoothness parameter lambda: 1e+06
## Number of iterations: 18
## n = 358, number of events = 266

We see that the effective dimension of the log-hazard is 2, so a straight line. Below is a plot of the hazard.

plot(ic_fit_gompertz) + ylim(0, 0.00125)

plot of chunk plot_gompertzplot of chunk plot_gompertz

The fact that the log-hazard is a straight line is easier to see after transforming the y-axis.

plot(ic_fit_gompertz) + scale_y_continuous(trans='log10')

plot of chunk plot_gompertz_logyplot of chunk plot_gompertz_logy

Note that the “Gompertz” hazard rate is implicitly obtained by setting the order of the penalty to 2. From the result of icfit the values of \(a\) and \(b\) and their standard errors can only be derived indirectly.

Weibull models

Also Weibull models can be fitted with icfit in the same indirect way, by setting bdeg, pord, and lambda, with one additional trick, namely time warping. The hazard of a Weibull distribution is of the form \(h(t) = \alpha \lambda t^{\alpha-1}\), with \(\lambda > 0\). If we take pord=2, lambda = 1e+06, and work with \(\log(t)\) instead of \(t\) itself, the form of the log-hazard would be \(\beta_0 + \beta_1 \log(t)\), which corresponds to a Weibull hazard with \(\alpha = \beta_1 + 1\), \(\lambda = e^{\beta_0} / (\beta_1 + 1)\).

Here is the fit of the Weibull model with icfit. This works fine with our present way of using time in days; time=1 corresponds to 1 day after randomisation. When time is in years, the logarithm of time would be negative for all events and censorings before one year. In that case adding a constant, or alternatively, defining logtime below as log(Ova$time * 365.25) would solve the issue.

Ova$logtime <- log(Ova$time)
ic_fit_weibull <- icfit(Surv(logtime, d) ~ 1, data = Ova,
                         bdeg=1, pord=2, lambda=1e+06, update_lambda=FALSE)
ic_fit_weibull
## Object of class 'icfit'
## Call:
## icfit(formula = Surv(logtime, d) ~ 1, data = Ova, lambda = 1e+06, 
##     bdeg = 1, pord = 2, update_lambda = FALSE)
## 
## Bins summary: tmin = 0, tmax = 7.991, number of bins = 100, bin width = 0.07991
## P-splines summary: number of segments = 20, degree = 1, penalty order = 2
## 
## Parameter estimates:
## 
## Effective baseline dimension: 1.047, log-likelihood: -1244, AIC: 2490
## Smoothness parameter lambda: 1e+06
## Number of iterations: 50
## n = 358, number of events = 266

Here is a plot of the hazard with log-transformed y-axis. Time on the x-axis actually refers to the logarithm (base 10) of the original time, so 3 on the x-axis corresponds to \(t=1000\).

plot(ic_fit_weibull) + scale_y_continuous(trans='log')

plot of chunk plot_icfit_weibullplot of chunk plot_icfit_weibull

To see everything on the original time scale and to facilitate future comparison with a direct parametric fit using survreg, let us extract the data elements of the plot, and explicitly transform time back into the original time.

We need to be careful though. Note that the estimated hazard with transformed time measures the expected number of events per transformed time unit, not in the original time unit. So we need not only back transform time, but also the hazard itself. If \(\tilde{h}(u)\) denotes the hazard in transformed time \(u = g(t)\), then in the original time scale \(t = f(u) = g^{-1}(u)\) we have that \(h(t) = \tilde{h}(u) / f^\prime(u)\). With \(u = g(t) = \log(t)\), \(t = f(u) = \exp(u)\), we get \(h(t) = \tilde{h}(u) / \exp(u) = \tilde{h}(\log(t)) / t\).

pw <- plot(ic_fit_weibull)$data

plot of chunk pweib

head(pw)[, 1:5]
##         time         haz        sehaz     lowerhaz    upperhaz
## 1 0.00000000 0.001471256 0.0004404592 0.0008181982 0.002645561
## 2 0.01598161 0.001492235 0.0004456058 0.0008311023 0.002679291
## 3 0.03196323 0.001513513 0.0004508096 0.0008442097 0.002713452
## 4 0.04794484 0.001535095 0.0004560715 0.0008575237 0.002748048
## 5 0.06392646 0.001556984 0.0004613920 0.0008710474 0.002783086
## 6 0.07990807 0.001579186 0.0004667716 0.0008847842 0.002818572
tail(pw)[, 1:5]
##         time      haz     sehaz lowerhaz upperhaz
## 496 7.910899 1.622971 0.1755261 1.312962 2.006176
## 497 7.926881 1.646086 0.1790758 1.329999 2.037294
## 498 7.942863 1.669530 0.1826939 1.347251 2.068901
## 499 7.958844 1.693308 0.1863816 1.364723 2.101006
## 500 7.974826 1.717424 0.1901403 1.382417 2.133616
## 501 7.990807 1.741885 0.1939712 1.400335 2.166739
pw$u <- pw$time
pw$time <- exp(pw$u)
pw$haz <- pw$haz / pw$time
pw$lowerhaz <- pw$lowerhaz / pw$time
pw$upperhaz <- pw$upperhaz / pw$time
head(pw)[, 1:5]
##       time         haz        sehaz     lowerhaz    upperhaz
## 1 1.000000 0.001471256 0.0004404592 0.0008181982 0.002645561
## 2 1.016110 0.001468576 0.0004456058 0.0008179255 0.002636812
## 3 1.032480 0.001465901 0.0004508096 0.0008176527 0.002628092
## 4 1.049113 0.001463231 0.0004560715 0.0008173799 0.002619402
## 5 1.066014 0.001460566 0.0004613920 0.0008171069 0.002610741
## 6 1.083187 0.001457906 0.0004667716 0.0008168338 0.002602109
tail(pw)[, 1:5]
##         time          haz     sehaz     lowerhaz     upperhaz
## 496 2726.842 0.0005951833 0.1755261 0.0004814956 0.0007357143
## 497 2770.771 0.0005940893 0.1790758 0.0004800103 0.0007352804
## 498 2815.408 0.0005929974 0.1826939 0.0004785278 0.0007348494
## 499 2860.765 0.0005919074 0.1863816 0.0004770483 0.0007344211
## 500 2906.851 0.0005908194 0.1901403 0.0004755718 0.0007339956
## 501 2953.681 0.0005897335 0.1939712 0.0004740984 0.0007335726

Here is a plot of the result.

plt <- ggplot(pw) +
  geom_ribbon(aes(x = pw$time, ymin = pw$lowerhaz, ymax = pw$upperhaz),
              fill = 'lightgrey') +
  geom_line(aes(x = pw$time, y = pw$lowerhaz), colour = 'blue') +
  geom_line(aes(x = pw$time, y = pw$upperhaz), colour = 'blue') +
  geom_line(aes(x = pw$time, y = pw$haz), size = 1, colour = 'blue', lty = 1) +
  ylim(0, 0.002) + 
  ggtitle("Weibull hazard fit") + xlab("Days since randomisation") +
  ylab("Hazard") + theme_bw() + theme(plot.title = element_text(hjust = 0.5))
plt

plot of chunk plot_pweib

Let us compare this now with survreg. It is most convenient to use the wrapper provided in flexsurvreg, because it uses a parametrisation close to ours. The parametrisation in flexsurvreg is of the form \(S(t) = \exp(-(t/\mu^\prime)^{\alpha^\prime})\), compared to our \(S(t) = \exp(-\lambda t^{\alpha})\), from which we see that \(\alpha = \alpha^\prime\) and \(\lambda = 1 / (\mu^\prime)^{\alpha^\prime}\). We add the fit from survreg with a dashed line to the original plot.

library(flexsurv)
fsr <- flexsurvreg(Surv(time, d) ~ 1, data = Ova, dist = "weibull")
fsr
## Call:
## flexsurvreg(formula = Surv(time, d) ~ 1, data = Ova, dist = "weibull")
## 
## Estimates: 
##        est       L95%      U95%      se      
## shape  9.20e-01  8.31e-01  1.02e+00  4.81e-02
## scale  1.32e+03  1.16e+03  1.51e+03  8.81e+01
## 
## N = 358,  Events: 266,  Censored: 92
## Total time at risk: 351135
## Log-likelihood = -2176.017, df = 2
## AIC = 4356.034
alp <- exp(fsr$coef[1])
mu <- exp(fsr$coef[2])
lam <- 1 / mu^alp
hw <- function(t, alp, lam) alp * lam * t^(alp - 1)
maxt <- max(Ova$time)
pw$haz_survreg <- hw(pw$time, alp, lam)

plt <- plt +
  geom_line(aes(x = pw$time, y = pw$haz_survreg), size = 1, lty = 2)
plt

plot of chunk flexsurvreg

The fit is quite similar, although not entirely equivalent.

References

Putter, H., Gampe, J., Eilers, P. H. (2023). Smooth hazard regression for interval censored data. Submitted for publication.