The AFT model framework

Estimation and inference

survreg

Accelerated Failure Time Models

Patrick Breheny

October 15

Patrick Breheny Survival Data Analysis (BIOS 7210) 1/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

Introduction

Last time, we introduced the Weibull distribution and saw

that, on the log scale, it could be viewed, essentially, as a
regular linear regression model, albeit with extreme value
residual terms
Today we will look at this modeling framework in more detail,
discuss estimation and model fitting, and go through an
example in which we fit the model to real data

Patrick Breheny Survival Data Analysis (BIOS 7210) 2/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

Accelerated failure time models

For a random time-to-event T , an accelerated failure time

(AFT) model proposes the following relationship between
covariates and Y = log T :

Yi = xTi β + Wi ,

iid
where Wi ∼ f are the error, or residual, terms; such models
are also sometimes referred to as log-linear models
The above framework describes a general class of models:
depending on the distribution we specify for W , we will obtain
a different model, but all will have the same general structure

Patrick Breheny Survival Data Analysis (BIOS 7210) 3/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

The Lognormal AFT

For example, an obvious possibility is to assume

iid
Wi ∼ N(0, σ 2 )
Assuming that Y follows a normal distribution is equivalent to
assuming that T follows a lognormal distribution
Thus, in the absence of censoring, we could simply use
ordinary least squares regression techniques to fit the model,
obtain confidence intervals, etc.

Patrick Breheny Survival Data Analysis (BIOS 7210) 4/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

The Lognormal AFT (cont’d)

Of course, we will almost always have some censoring, so we

will have to extend ordinary linear model methods to handle
censoring
Furthermore, the lognormal distribution, while convenient,
does not accurately describe most time-to-event distributions:
0.30

0.25

0.20
λ(t)

0.15

0.10

0.05

0 5 10 15 20

Patrick Breheny Survival Data Analysis (BIOS 7210) 5/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

AFTs rescale time

For any AFT, we have

T = eηi T0 ,

where T0 = eW and ηi = xTi β

In other words, whereas in a proportional hazards (PH) model,
the covariates act multiplicatively on the hazard, in an AFT
model the covariates act multiplicatively on time:
If eηi = 1/2, that subject effectively ages at twice normal speed
If eηi = 2, that subject effectively ages at only half of normal
speed

Patrick Breheny Survival Data Analysis (BIOS 7210) 6/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

Survival

Si (t) = S0 (e−ηi t)
Baseline eηi = 2 eηi = 1 2

1.0

0.8

0.6
S(t)

0.4

0.2

0.0

0.0 0.2 0.4 0.6 0.8 1.0

Patrick Breheny Survival Data Analysis (BIOS 7210) 7/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

Hazard

λi (t) = λ0 (e−ηi t)e−ηi

Baseline e−ηi = 3 2 e−ηi = 2 3

3.0

2.5

2.0
λ(t)

1.5

1.0

0.5

0.0

0.0 0.2 0.4 0.6 0.8 1.0

Patrick Breheny Survival Data Analysis (BIOS 7210) 8/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

AFT vs PH: Hazard (eηi = 2/3 for PH; eηi = 3/2 for AFT)

PH : λi (t) = λ0 (t)eηi
AFT : λi (t) = λ0 (e−ηi t)e−ηi
Baseline PH AFT

2.0

1.5
λ(t)

1.0

0.5

0.0

0.0 0.2 0.4 0.6 0.8 1.0

Patrick Breheny Survival Data Analysis (BIOS 7210) 9/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

AFT vs PH: Log scale (eηi = 2/3 for PH, 3/2 for AFT)

PH : λi (y) = λ̃0 (y)eηi

AFT : λi (y) = λ̃0 (y − ηi )
Baseline PH AFT

0.5

0.0

−0.5
log(λ(y))

−1.0

−1.5

−2.0

−2.5

−2.0 −1.5 −1.0 −0.5 0.0 0.5

Patrick Breheny Survival Data Analysis (BIOS 7210) 10/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

Remarks

As the final plot depicts, on the log-time vs. log-hazard scale,

the effect of the PH assumption is to model the change in
hazard as a vertical shift, while the effect of the AFT
assumption is to model the change in hazard as a horizontal
shift
In general, the two assumptions cannot be reconciled – on the
previous plot, there is no way to shift the gray line vertically
and obtain the blue line, for example – and therefore, a model
is typically either an AFT model or a PH model, but not both

Patrick Breheny Survival Data Analysis (BIOS 7210) 11/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

Weibull regression satisfies both AFT and PH

There is one exception, however: if the distribution is linear

(on the log-time vs. log-hazard scale), then any vertical shift
of the line will correspond to a horizontal shift
Recall that the extreme value distribution, λ(y) = ey , is linear
on this scale, and that the Weibull distribution represents its
location-scale family (i.e., all possible intercepts and slopes)
Thus, the Weibull distribution is the unique distribution
(along with its special cases, like the exponential distribution)
that satisfies both the PH and AFT assumptions (see section
2.3.4 of our book for a more formal proof)

Patrick Breheny Survival Data Analysis (BIOS 7210) 12/25

 The AFT model framework Introduction
Estimation and inference Example: The lognormal AFT
survreg Meaning of AFT models

Interpretation of coefficients

Finally, let us briefly consider the interpretation of the

coefficients in an AFT model
As we have seen, the effect of a δj -unit change in covariate j
is to multiply the failure time by exp(δj βj )
For example, if xj was a treatment indicator and βj = 0.4, we
could say that patients who received the treatment survived
50% longer than patients who did not receive the treatment

Patrick Breheny Survival Data Analysis (BIOS 7210) 13/25

 The AFT model framework
Estimation
Estimation and inference
Inference
survreg

Maximum likelihood estimation: Setup

Suppose that we can write our AFT model

Yi = xTi β + σWi ;

note, for example, that the Weibull and lognormal models can
be written this way
Then the likelihood is
Y
L(β, σ|y, d) = {σ −1 f (wi )}di {S(wi )}1−di
i
Y
= {σ −1 λ(wi )}di S(wi ),
i

where f , λ, and S represent the density, hazard, and survival

functions for the error distribution, and wi = (yi − xTi β)/σ
Patrick Breheny Survival Data Analysis (BIOS 7210) 14/25
 The AFT model framework
Estimation
Estimation and inference
Inference
survreg

Score function(s)

For the most part, estimation proceeds as it did in the

exponential regression example we discussed previously, except
that we now have to estimate a scale parameter σ as well
So, for example,
∂`
= −σ −1 XT a
∂β
∂`
= −σ −1 (d + wT a),
∂σ
P
where d = i di and ai = ∂`i /∂wi
For example, ai = di − ewi when W follows an extreme value
distribution

Patrick Breheny Survival Data Analysis (BIOS 7210) 15/25

 The AFT model framework
Estimation
Estimation and inference
Inference
survreg

Newton-Raphson algorithm

Generally speaking, a will be a nonlinear function of β and σ,

meaning that to solve u(β, σ) = 0 we again must use the
Newton-Raphson approach we discussed previously
One complication here is that it is possible to obtain
unreasonable updates for σ (e.g., negative values)
There are various ways to deal with this issue; one is to
update σ more gradually by

σm+1 = (1 − τ )σm + τ σ̃,

where σ̃ is the Newton-Raphson update

Patrick Breheny Survival Data Analysis (BIOS 7210) 16/25

 The AFT model framework
Estimation
Estimation and inference
Inference
survreg

Information matrix

Again, the Wald approach is the most convenient, although

not the most accurate, for inference
To proceed, we need (the inverse of) the information matrix:
" 2 2
#
∂ ` ∂ `
∂β 2 ∂β∂σ
I(θ̂) = − ∂2` ∂2`
;
∂σ∂β ∂σ 2

expressions for these quantities are given in the book

It is worth noting that the Wald approach is typically more
accurate if the likelihood is parameterized in terms of
τ = log σ; this is the approach taken in the survival package

Patrick Breheny Survival Data Analysis (BIOS 7210) 17/25

 The AFT model framework
Estimation
Estimation and inference
Inference
survreg

Accounting for uncertainty in σ

As we remarked previously, it is important to note that

−1
∂2`


∂β 2

underestimates the uncertainty with respect to β

Instead, we must consider the upper left p × p block of I−1 ,
which takes into account the fact that uncertainty about σ
increases standard errors for β (and vice versa)

Patrick Breheny Survival Data Analysis (BIOS 7210) 18/25

 The AFT model framework
Estimation and inference
survreg

Basic usage

The survival package offers a function, survreg, for fitting

parametric AFT models
The syntax is similar to other regression modeling functions in
R:
survreg(S ~ trt + stage + hepato + bili, pbc)

where S is a Surv object

The default is to use a Weibull distribution, but exponential,
lognormal, and other distributions are available using the
dist= option

Patrick Breheny Survival Data Analysis (BIOS 7210) 19/25

 The AFT model framework
Estimation and inference
survreg

Exponential regression

Thus, for example,

survreg(S ~ trt + stage + hepato + bili, pbc,
dist='exponential')

Note that coef(fit) exactly matches our do-it-yourself

regression results from last week, except that all the signs are
reversed (increasing hazard means decreasing failure time)
Furthermore, vcov(fit) exactly matches the inverse of our
information matrix from last week

Patrick Breheny Survival Data Analysis (BIOS 7210) 20/25

 The AFT model framework
Estimation and inference
survreg

Weibull results

> fit <- survreg(S ~ trt + stage + hepato + bili, pbc)

> summary(fit)

Value Std. Error z p

(Intercept) 9.8086 0.34509 28.423 1.05e-177
trt 0.0960 0.11510 0.834 4.04e-01
stage -0.4294 0.09045 -4.747 2.06e-06
hepato -0.2363 0.14625 -1.616 1.06e-01
bili -0.0915 0.00923 -9.916 3.53e-23
Log(scale) -0.3760 0.06898 -5.451 5.00e-08

Scale= 0.687
Number of Newton-Raphson Iterations: 6

Patrick Breheny Survival Data Analysis (BIOS 7210) 21/25

 The AFT model framework
Estimation and inference
survreg

Remarks

Broadly speaking, these results are similar to what we obtained

with exponential regression (bilirubin and stage clearly
significant, treatment and hepatomegaly not significant)
The coefficients are somewhat different, of course, and have a
different interpretation: a one-unit change in stage shortens
survival time by 35% (e−0.4294 = 0.65)
Also, note that the summary provides a test of whether
log σ = 0; in other words, for the adequacy of exponential
regression, which is firmly rejected here despite the diagnostic
plot looking OK

Patrick Breheny Survival Data Analysis (BIOS 7210) 22/25

 The AFT model framework
Estimation and inference
survreg

Weibull proportional hazards model

The survival package does not offer a function for fitting

(parametric) proportional hazards models
Recall, however, that exponential and Weibull proportional
hazards models can be reparameterized as AFT models
Thus, we can obtain the PH coefficients by fitting the AFT
model and using the transformation β = −β ∗ /σ, where β
denotes the PH coefficients and β ∗ denotes the AFT
coefficients:
b.ph <- -coef(fit.aft)/fit.aft$scale

Patrick Breheny Survival Data Analysis (BIOS 7210) 23/25

 The AFT model framework
Estimation and inference
survreg

Comparison of estimates

AFT PH
Exponential Weibull Exponential Weibull
trt 0.14 0.10 -0.14 -0.14
stage -0.56 -0.43 0.56 0.63
hepato -0.34 -0.24 0.34 0.34
bili -0.11 -0.09 0.11 0.13

Patrick Breheny Survival Data Analysis (BIOS 7210) 24/25

 The AFT model framework
Estimation and inference
survreg

Predictions: Stage 3
Predicted survival curves for a patient on the placebo arm, stage 3
cirrhosis, no hepatomegaly, and bili=1
Exponential Weibull

1.0
Progression−free survival

0.8

0.6

0.4

0.2

0.0

0 2 4 6 8 10

Time (years)

Patrick Breheny Survival Data Analysis (BIOS 7210) 25/25