1 Introduction
In a comparative, longitudinal clinical study, often the primary
endpoint is the time to a specific clinical event, such as death, heart
failure hospitalization, tumor progression, and so on. The hazard ratio
estimate is almost routinely used to quantify the treatment difference.
However, the clinical meaning of such a model-based between-group
summary can be rather difficult to interpret when the underlying model
assumption (i.e., the proportional hazards assumption) is violated, and
it is difficult to assure that the modeling is indeed correct
empirically. For example, a non-significant result of a goodness-of-fit
test does not necessary mean that the proportional hazards assumption is
“correct.” Other issues on the hazard ratio is seen elsewhere [1, 2].
Between-group summery metrics based on the restricted mean survival time
(RMST) are useful alternatives to the hazard ratio or other model-based
measures. This vignette is a supplemental documentation for
survRM2 package and illustrates how to use the functions in the
package to compare two groups with respect to the restricted mean
survival time. The package was made and tested on R version 3.3.2.
2 Sample data
Throughout this vignette, we use a part of data from the primary
biliary cirrhosis (pbc) study conducted by the Mayo Clinic, which is
included in survival package in R. The details of the study and the data
elements are seen in the help file in survival package, which
can be seen by
The original data in the survival package consists of data from 418
patients, which includes those who participated in the randomized
clinical trial and those who did not. In the following illustration, we
use only 312 cases who participated in the randomized trial (158 cases
on D-penicillamine group and 154 cases on Placebo group). The following
function in survRM2 package creates the data used in this
vignette, selecting the subset from the original data file.
library(survRM2)
D = rmst2.sample.data()
nrow(D)
## [1] 312
## time status arm
## 1 1.095140 1 1
## 2 12.320329 0 1
## 3 2.770705 1 1
## 4 5.270363 1 1
## 5 4.117728 0 0
## 6 6.852841 1 0
Here, time is years from the registration to death
or last known alive, status is the indicator of the
event (1: death, 0: censor), and arm is the treatment
assignment indicator (1: D-penicillamine, 0: Placebo). Below is the
Kaplan-Meier (KM) estimate for time-to-death of each treatment
group.
3 Restricted mean survival time (RMST) and restricted mean time lost
(RMTL)
The RMST is defined as the area under the curve of the survival
function up to a time \(\tau (<
\infty):\) \[ \mu_{\tau} =
\int_0^{\tau} S(t)dt,\] where \(S(t)\) is the survival function of a
time-to-event variable of interest. The interpretation of the RMST is
that “when we follow up patients for \(\tau,\) patients will survive for \(\mu_{\tau}\) on average,” which is quite
straightforward and clinically meaningful summary of the censored
survival data. If there were no censored observations, one could use the
mean survival time \[ \mu_{\infty} =
\int_0^{\infty} S(t)dt,\] instead of \(\mu_{\tau}.\)
A natural estimator for \(\mu_{\tau}\) is \[ \hat{\mu}_{\tau} = \int_0^{\tau}
\hat{S}(t)dt,\] where \(\hat{S}(t)\) is the KM estimator for \(S(t).\) The standard error for \(\hat{\mu}_{\tau}\) is also calculated
analytically; the detailed formula is given in [3]. Note that \(\mu_{\tau}\) is estimable even under a
heavy censoring case. On the other hand, although median survival time,
\(S^{-1}(0.5),\) is also a robust
summary of survival time distribution, it will become inestimable when
the KM curve does not reach 0.5 due to heavy censoring or rare
events.
The RMTL is defined as the area “above” the curve of the survival
function up to a time \(\tau:\) \[ \tau - \mu_{\tau} = \int_0^{\tau} \{ 1-S(t)
\}dt.\] In the following figure, the area highlighted in pink and
orange are the RMST and RMTL estimates, respectively, in D-penicillamine
group, when \(\tau\) is 10 years. The
result shows that the average survival time during 10 years of follow-up
is 7.15 years in the D-penicillamine group. In other words, during the
10 years of follow-up, patients treated by D-penicillamine lost 2.85
years in average sense.
3.1 Unadjusted analysis and its implementation
Let \(\mu_{\tau}(1)\) and \(\mu_{\tau}(0)\) denote the RMST for
treatment group 1 and 0, respectively. Now, we compare the two survival
curves, using the RMST or RMTL. Specifically, we consider the following
three measures for the between-group contrast.
- Difference in RMST \[ \mu_{\tau}(1) -
\mu_{\tau}(0) \]
- Ratio of RMST \[ \mu_{\tau}(1) /
\mu_{\tau}(0) \]
- Ratio of RMTL \[ \{ \tau - \mu_{\tau}(1)
\} / \{ \tau - \mu_{\tau}(0) \} \]
These are estimated by simply replacing \(\mu_{\tau}(1)\) and \(\mu_{\tau}(0)\) by their empirical
counterparts (i.e.,\(\hat{\mu}_{\tau}(1)\) and \(\hat{\mu}_{\tau}(0)\), respectively). For
inference of the ratio type metrics, we use the delta method to
calculate the standard error. Specifically, we consider \(\log \{ \hat{\mu}_{\tau}(1) \}\) and \(\log \{ \hat{\mu}_{\tau}(0) \}\) and
calculate the standard error of log-RMST. We then calculate a confidence
interval for log-ratio of RMST, and transform it back to the original
ratio scale. Below shows how to use the function,
rmst2, to implement these analyses.
time = D$time
status = D$status
arm = D$arm
rmst2(time, status, arm, tau=10)
The first argument (time) is the time-to-event
vector variable. The second argument (status) is also a
vector variable with the same length as time, each of
the elements takes either 1 (if event) or 0 (if no event). The third
argument (arm) is a vector variable to indicate the
assigned treatment of each subject; the elements of this vector take
either 1 (if active treatment arm) or 0 (if control arm). The fourth
argument (tau) is a scalar value to specify the
truncation time point \({\bf \tau}\)
for the RMST calculation. Note that \(\tau\) needs to be smaller than the minimum
of the largest observed time in each of the two groups (let us call this
the max \(\tau\)). The program will
stop with an error message when such \(\tau\) is specified. When \(\tau\) is not specified in
rmst2, i.e., when the code looks like
the max \(\tau\) is used as the
default \(\tau.\) It is always
encouraged to confirm that the size of the risk set is large enough at
the specified \(\tau\) in each group to
make sure the stability of the KM estimates.
Below is the output with the pbc example when \(\tau=10\) (years) is specified. The
rmst2 function returns RMST and RMTL on each group and
the results of the between-group contrast measures listed above.
obj = rmst2(time, status, arm, tau=10)
print(obj)
##
## The truncation time: tau = 10 was specified.
##
## Restricted Mean Survival Time (RMST) by arm
## Est. se lower .95 upper .95
## RMST (arm=1) 7.146 0.283 6.592 7.701
## RMST (arm=0) 7.283 0.295 6.704 7.863
##
##
## Restricted Mean Time Lost (RMTL) by arm
## Est. se lower .95 upper .95
## RMTL (arm=1) 2.854 0.283 2.299 3.408
## RMTL (arm=0) 2.717 0.295 2.137 3.296
##
##
## Between-group contrast
## Est. lower .95 upper .95 p
## RMST (arm=1)-(arm=0) -0.137 -0.939 0.665 0.738
## RMST (arm=1)/(arm=0) 0.981 0.878 1.096 0.738
## RMTL (arm=1)/(arm=0) 1.050 0.787 1.402 0.738
In the present case, the difference in RMST (the first row of the
“Between-group contrast” block in the output) was -0.137 years. The
point estimate indicated that patients on the active treatment survive
0.137 years shorter than those on placebo group on average, when
following up the patients 10 years. While no statistical significance
was observed (p=0.738), the 0.95 confidence interval (-0.665 to 0.939)
was relatively tight around 0, suggesting that the difference in RMST
would be at most +/- one year.
The package also has a function to generate a plot from the
rmst2 object. The following figure is automatically
generated by simply passing the resulting rmst2 object
to plot() function after running the aforementioned unadjusted
analyses.
plot(obj, xlab="Years", ylab="Probability")
3.2 Adjusted analysis and implementation
In most of the randomized clinical trials, an adjusted analysis is
usually included in one of the planned analyses. One reason would be
that adjusting for important prognostic factors may increase power to
detect a between-group difference. Another reason would be we sometimes
observe imbalance in distribution of some of baseline prognostic factors
even though the randomization guarantees the comparability of the two
groups on average. The function, rmst2, in this package
implements an ANCOVA type adjusted analysis proposed by Tian et al. [4],
in addition to the unadjusted analyses presented in the previous
section.
Let \(Y\) be the restricted mean
survival time, and let \(Z\) be the
treatment indicator. Also, let \(X\)
denote a \(q\)-dimensional baseline
covariate vector. Tian’s method consider the following regression model,
\[ g\{ E(Y \mid Z, X) \} = \alpha + \beta Z +
\gamma^\prime X, \] where \(g(\cdot)\) is a given smooth and strictly
increasing link function, and \((\alpha,
\beta, \gamma^\prime)\) is a \((q+2)\)-dimension unknown parameter vector.
Prior to Tian et al. [4], Andersen et al. [5] also studied this
regression model and proposed an inference procedure for the unknown
model parameter, using a pseudo-value technique to handle censored
observations. Program codes for their pseudo-value approach are
available on the three major platforms (Stata, R and SAS) with detailed
documentation [6, 7]. In contrast to Andersen’s method [5, 6, 7], Tian’s
method [4] utilizes an inverse probability censoring weighting technique
to handle censored observations. The function, rmst2,
in this package implements this method.
As shown below, for implementation of Tian’s adjusted analysis for
the RMST, the only the difference is if the user passes covariate data
to the function. Below is a sample code to perform the adjusted
analyses.
rmst2(time, status, arm, tau=10, covariates=x)
where covariates is the argument for a vector/matrix
of the baseline characteristic data, x. For illustration, let us try the
following three baseline variables, in the pbc data, as the covariates
for adjustment.
## age bili albumin
## 1 58.76523 14.5 2.60
## 2 56.44627 1.1 4.14
## 3 70.07255 1.4 3.48
## 4 54.74059 1.8 2.54
## 5 38.10541 3.4 3.53
## 6 66.25873 0.8 3.98
The rmst2 function fits data to a model for each of
the three contrast measures (i.e., difference in RMST, ratio of RMST,
and ratio of RMTL). For the difference metric, the link function \(g(\cdot)\) in the model above is the
identity link. For the ratio metrics, the log-link is used.
Specifically, with this pbc example, we are now trying to fit data to
the following regression models:
- Difference in RMST \[ E(Y \mid arm,\ X) =
\alpha + \beta (arm) + \gamma_1 (age) + \gamma_2(bili) +
\gamma_3(albumin), \]
- Ratio of RMST \[ \log \{ E(Y \mid arm, \
X) \} = \alpha + \beta (arm) + \gamma_1 (age) + \gamma_2(bili) +
\gamma_3(albumin), \]
- Ratio of RMTL \[ \log \{ \tau - E(Y \mid
arm, \ X) \} = \alpha + \beta (arm) + \gamma_1 (age) + \gamma_2(bili) +
\gamma_3(albumin). \]
Below is the output that rmst2 returns for the
adjusted analyses.
rmst2(time, status, arm, tau=10, covariates=x)
##
## The truncation time: tau = 10 was specified.
##
## Summary of between-group contrast (adjusted for the covariates)
## Est. lower .95 upper .95 p
## RMST (arm=1)-(arm=0) -0.210 -0.883 0.463 0.540
## RMST (arm=1)/(arm=0) 0.968 0.877 1.068 0.514
## RMTL (arm=1)/(arm=0) 1.035 0.806 1.329 0.786
##
##
## Model summary (difference of RMST)
## coef se(coef) z p lower .95 upper .95
## intercept 2.743 2.134 1.285 0.199 -1.440 6.927
## arm -0.210 0.343 -0.613 0.540 -0.883 0.463
## age -0.069 0.018 -3.900 0.000 -0.103 -0.034
## bili -0.325 0.039 -8.386 0.000 -0.401 -0.249
## albumin 2.550 0.472 5.401 0.000 1.624 3.475
##
##
## Model summary (ratio of RMST)
## coef se(coef) z p exp(coef) lower .95 upper .95
## intercept 1.369 0.356 3.842 0.000 3.930 1.955 7.899
## arm -0.033 0.050 -0.652 0.514 0.968 0.877 1.068
## age -0.009 0.003 -3.410 0.001 0.991 0.985 0.996
## bili -0.087 0.013 -6.523 0.000 0.917 0.893 0.941
## albumin 0.360 0.080 4.491 0.000 1.434 1.225 1.678
##
##
## Model summary (ratio of time-lost)
## coef se(coef) z p exp(coef) lower .95 upper .95
## intercept 1.992 0.695 2.865 0.004 7.332 1.876 28.655
## arm 0.035 0.127 0.272 0.786 1.035 0.806 1.329
## age 0.025 0.007 3.810 0.000 1.026 1.012 1.039
## bili 0.063 0.008 8.334 0.000 1.065 1.049 1.080
## albumin -0.750 0.149 -5.033 0.000 0.472 0.353 0.633
The first block of the output is a summary of the adjusted treatment
effect. Subsequently, a summary for each of the three models are
provided.
References
[1] Hernan, M. A. (2010). The hazards of hazard ratios.
Epidemiology (Cambridge, Mass) 21, 13-15.
[2] Uno, H., Claggett, B., Tian, L., Inoue, E., Gallo, P., Miyata,
T., Schrag, D., Takeuchi, M., Uyama, Y., Zhao, L., Skali, H., Solomon,
S., Jacobus, S., Hughes, M., Packer, M. & Wei, L.-J. (2014). Moving
beyond the hazard ratio in quantifying the between-group difference in
survival analysis. Journal of clinical oncology : official journal
of the American Society of Clinical Oncology 32,
2380-2385.
[3] Miller, R. G. (1981). Survival Analysis. Wiley.
[4] Tian, L., Zhao, L. & Wei, L. J. (2014). Predicting the
restricted mean event time with the subject’s baseline covariates in
survival analysis. Biostatistics 15,
222-233.
[5] Andersen, P. K., Hansen, M. G. & Klein, J. P. (2004).
Regression analysis of restricted mean survival time based on
pseudo-observations. Lifetime data analysis
10, 335-350.
[6] Klein, J. P., Gerster, M., Andersen, P. K., Tarima, S. &
Perme, M. P. (2008). SAS and R functions to compute pseudo-values for
censored data regression. Computer methods and programs in
biomedicine 89, 289-300.
[7] Parner, E. T. & Andersen, P. K. (2010). Regression analysis
of censored data using pseudo-observations. The Stata Journal
10(3), 408-422.
These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.