DISTRIBUTION PARAMETERIZATION FROM SURVIVAL ANALYSIS in R
This document is a guide to interpret distribution parameters obtained from R survival analysis
using SURVREG and FLEXSURVREG packages and to use them properly within TreeAge Pro’s
distributions parameters.
Please note that SURVREG and FLEXSURVREG generate parameters which are presented differently and
often need to be further transformed with exp( ) or log( ) expressions in order to be equivalent
to each other.
In the examples below a following steps were performed in R.
1. Generated 100,000 samples from a particular distribution with given input parameters.
2. Feed the 100,000 samples into SURVREG and FLEXSURVREG (no censoring) to obtain the
estimates of the parameters for the given distribution.
3. The estimated parameters must match the input parameters from step 1, but often need to
be transformed.
a. Some distributions have different parameterization in TreeAge Pro and in R,
appropriate transformation of parameters for TreeAge Pro (TP) are shown in square
green shaded boxes.
LOGNORMAL DISTRIBUTION
LOGNORMAL SAMPLES from R function:
Y = rlnorm(100000,meanlog = 1.1, sdlog = 1.2)
The Results from the SURVREG function
Call:
survreg(formula = Surv(time, psurv) ~ 1, data = myData2, dist = "lognormal")
Coefficients:
(Intercept) Mu (Mean of Logs) = 1.10127
1.10127
Scale= 1.201226 Sigma (std. dev. of logs) = 1.201226
Loglik(model)= -270355.1 Loglik(intercept only)= -270355.1
n= 100000
********************************
The Results from the FLEXSURVREG function
Call:
flexsurvreg(formula = Surv(time, psurv) ~ 1, data = myData2,
dist = "lognormal")
Estimates: Mu (Mean of Logs) = 1.10127
est L95% U95% se
meanlog 1.10127 1.09383 1.10872 0.00380
sdlog 1.20123 1.19597 1.20650 0.00269
Sigma (std. dev. of logs) = 1.20123
N = 100000, Events: 100000, Censored: 0
Total time at risk: 617840.2
Log-likelihood = -270355.1, df = 2
AIC = 540714.3
*********************
Mean of Samples = 6.178402, Standard Deviation = 10.883834
TP Parameters mu (Mean of Logs) = 1.1 sigma (std. dev. of logs) = 1.2
LOGLOGISTIC DISTRIBUTION - notice the confusing implementation of R parameterizations of SURVREG output!
********************************
LOGLOGISTIC SAMPLES from R function:
Y = rllogis(100000, shape = 1.5, scale = 1.2)
The Results from the SURVREG function
Call:
survreg(formula = Surv(time, psurv) ~ 1, data = myData2, dist = "loglogistic")
Coefficients:
(Intercept) a = exp(0.181) ≈ 1.2
0.1814073
Scale= 0.6679469 b = 1/0.6679 ≈ 1.5
Loglik(model)= -177722.6 Loglik(intercept only)= -177722.6
n= 100000
********************************
The Results from the FLEXSURVREG function
Call:
flexsurvreg(formula = Surv(time, psurv) ~ 1, data = myData2,
dist = "llogis")
Estimates:
est L95% U95% se b ≈ 1.5
shape 1.49706 1.48932 1.50483 0.00396
scale 1.19889 1.19032 1.20752 0.00439
a ≈ 1.2
N = 100000, Events: 100000, Censored: 0
Total time at risk: 301627.5
Log-likelihood = -177722.6, df = 2
AIC = 355449.3
*********************
Mean of Samples = 3.016275, Standard Deviation = 58.476523
TP Parameters a = 1.2 b = 1.5
WEIBULL DISTRIBUTION - notice the confusing implementation of R parameterizations of SURVREG output!
********************************
WEIBULL SAMPLES from R function:
Y = rweibull(100000, shape = 1.5, scale = 1.2)
The Results from the SURVREG function
Call:
survreg(formula = Surv(time, psurv) ~ 1, data = myData2, dist = "weibull")
Coefficients:
(Intercept) Use alternate parameter λ_w = exp(0.181) ≈ 1.2
0.1831886
Scale= 0.6642631 k = 1/0.6679 ≈ 1.5
Loglik(model)= -96781.6 Loglik(intercept only)= -96781.6
n= 100000
********************************
The Results from the FLEXSURVREG function
Call:
flexsurvreg(formula = Surv(time, psurv) ~ 1, data = myData2,
dist = "weibull")
Estimates:
est L95% U95% se k ≈ 1.5
shape 1.50543 1.49817 1.51272 0.00371
scale 1.20104 1.19585 1.20626 0.00266
Use alternate parameter λ_w ≈ 1.2
N = 100000, Events: 100000, Censored: 0
Total time at risk: 108371.7
Log-likelihood = -96781.64, df = 2
AIC = 193567.3
*********************
Mean of Samples = 1.083717, Standard Deviation = 0.733614
TP Parameters Alternative Parameters λ_w = 1.2 k = 1.5
EXPONENTIAL DISTRIBUTION - notice the confusing implementation of R parameterizations of SURVREG output!
********************************
EXPONENTIAL SAMPLES from R function:
Y = rexp(100000, rate = 1.2)
The Results from the SURVREG function
Call:
survreg(formula = Surv(time, psurv) ~ 1, data = myData2, dist = "exponential")
Coefficients:
(Intercept) λ = exp(- (-0.181) ) ≈ 1.2
-0.1873548
Scale fixed at 1
Loglik(model)= -81264.5 Loglik(intercept only)= -81264.5
n= 100000
********************************
The Results from the FLEXSURVREG function
Call:
flexsurvreg(formula = Surv(time, psurv) ~ 1, data = myData2,
dist = "exp")
Estimates:
est L95% U95% se
rate 1.20606 1.19860 1.21355 0.00381
λ ≈ 1.2
N = 100000, Events: 100000, Censored: 0
Total time at risk: 82914.95
Log-likelihood = -81264.52, df = 1
AIC = 162531
*********************
Mean of Samples = 0.829149, Standard Deviation = 0.828808
TP Parameter λ = 1.2
GENERALIZED GAMMA DISTRIBUTION - not supported by R SURVREG package.
********************************
GENERALIZED GAMMA (ORIGINAL) SAMPLES from R function:
Y = rgengamma.orig(100000, shape = 1.2, scale=0.9, k=1.3)
The Results from the FLEXSURVREG function
Call:
flexsurvreg(formula = Surv(time, psurv) ~ 1, data = myData2,
dist = "gengamma.orig")
Estimates:
est L95% U95% se c ≈ 1.2
shape 1.1867 1.1616 1.2123 0.0129
scale 0.8879 0.8522 0.9250 0.0186
Beta ≈ 0.9
k 1.3211 1.2750 1.3690 0.0240
N = 100000, Events: 100000, Censored: 0 Alpha ≈ 1.3
Total time at risk: 107271
Log-likelihood = -98912.74, df = 3
AIC = 197831.5
*********************
Mean of Samples = 1.063595, Standard Deviation = 0.7829048
TP Parameters c = 1.2, Alpha =1.3, Beta = 0.9
