Empirical Bayes Inference in Structured Hazard Regression...Empirical Bayes Inference in Structured...
Transcript of Empirical Bayes Inference in Structured Hazard Regression...Empirical Bayes Inference in Structured...
Empirical Bayes Inference inStructured Hazard Regression
Thomas Kneib
Department of StatisticsLudwig-Maximilians-University Munich
20.11.2006
Thomas Kneib Outline
Outline
• Leukemia survival data.
• Structured hazard regression for continuous survival times.
• Empirical Bayes inference in structured hazard regression.
• Multi-state models.
Empirical Bayes Inference in Structured Hazard Regression 1
Thomas Kneib Leukemia Survival Data
Leukemia Survival Data
• Survival times of adults after diagnosis of acute myeloid leukemia.
• 1,043 cases diagnosed between 1982 and 1998 in Northwest England.
• 16 % (right) censored.
• Continuous and categorical covariates:
age age at diagnosis,wbc white blood cell count at diagnosis,sex sex of the patient,tpi Townsend deprivation index.
• Spatial information in different resolution.
Empirical Bayes Inference in Structured Hazard Regression 2
Thomas Kneib Leukemia Survival Data
• Classical Cox proportional hazards model:
λ(t; x) = λ0(t) exp(x′γ).
• Baseline hazard λ0(t) is a nuisance parameter and remains unspecified.
• Estimate γ based on the partial likelihood.
• Questions / Limitations:
– Simultaneous estimation of baseline hazard rate and covariate effects.
– Flexible modelling of covariate effects (e.g. nonlinear effects, interactions).
– Spatially correlated survival times.
– Non-proportional hazards models / time-varying effects.
⇒ Structured hazard regression models.
Empirical Bayes Inference in Structured Hazard Regression 3
Thomas Kneib Leukemia Survival Data
• Replace usual parametric predictor with a flexible semiparametric predictor
λ(t; ·) = λ0(t) exp[g1(t)sex + f1(age) + f2(wbc) + f3(tpi) + fspat(si)]
and absorb the baseline
λ(t; ·) = exp[g0(t) + g1(t)sex + f1(age) + f2(wbc) + f3(tpi) + fspat(si)]
where
– g0(t) = log(λ0(t)) is the log-baseline hazard,
– g1(t) is a time-varing gender effect,
– f1, f2, f3 are nonparametric functions of age, white blood cell count anddeprivation, and
– fspat is a spatial function.
Empirical Bayes Inference in Structured Hazard Regression 4
Thomas Kneib Leukemia Survival Data
−5
−2
14
0 7 14time in years
log(baseline) Log-baseline hazard.
Time-varying gender effect.
−1
01
2
0 7 14time in years
time−varying effect of sex
Empirical Bayes Inference in Structured Hazard Regression 5
Thomas Kneib Leukemia Survival Data
−1.5
−.7
50
.75
1.5
14 27 40 53 66 79 92age in years
effect of age Effect of age at diagnosis.
Effect of white blood cell count.−
.50
.51
1.5
2
0 125 250 375 500
white blood cell count
effect of white blood cell count
Empirical Bayes Inference in Structured Hazard Regression 6
Thomas Kneib Leukemia Survival Data
−.5
−.2
50
.25
.5
−6 −2 2 6 10townsend deprivation index
effect of townsend deprivation index Effect of deprivation.
Empirical Bayes Inference in Structured Hazard Regression 7
Thomas Kneib Leukemia Survival Data
-0.44 0 0.3
District-level analysis−0.44 0.3
Individual-level analysis
Empirical Bayes Inference in Structured Hazard Regression 8
Thomas Kneib Structured Hazard Regression
Structured Hazard Regression
• A general structured hazard regression model consists of an arbitrary combination ofthe following model terms:
– Log baseline hazard g0(t) = log(λ0(t)).
– Time-varying effects gl(t)ul of covariates ul.
– Nonparametric effects fj(xj) of continuous covariates xj.
– Spatial effects fspat(s) of a spatial location variable s.
– Interaction surfaces fj,k(xj, xk) of two continuous covariates.
– Varying coefficient interactions ujfk(xk) or ujfspat(s).
– Frailty terms bg (random intercept) or xjbg (random slopes).
• All covariates are themselves allowed to be (piecewise constant) time-varying.
Empirical Bayes Inference in Structured Hazard Regression 9
Thomas Kneib Structured Hazard Regression
• Penalised splines for the baseline effect, time-varying effects, and nonparametriceffects:
– Approximate f(x) (or g(t)) by a weighted sum of B-spline basis functions
f(x) =∑
ξjBj(x).
– Employ a large number of basis functions to enable flexibility.
– Penalise differences between parameters of adjacent basis functions to ensuresmoothness:
Pen(ξ|τ2) =1
2τ2
∑(∆kξj)2.
– Bayesian interpretation: Assume a k-th order random walk prior for ξj, e.g.
ξj = ξj−1 + uj, uj ∼ N(0, τ2) (RW1).
ξj = 2ξj−1 − ξj−2 + uj, uj ∼ N(0, τ2) (RW2).
Empirical Bayes Inference in Structured Hazard Regression 10
Thomas Kneib Structured Hazard Regression
−2
−1
01
2
−3 −1.5 0 1.5 3
−2
−1
01
2
−3 −1.5 0 1.5 3
−2
−1
01
2
−3 −1.5 0 1.5 3
Empirical Bayes Inference in Structured Hazard Regression 11
Thomas Kneib Structured Hazard Regression
• Bivariate Tensor product P-splines for interaction surfaces:
– Define bivariate basis functions (Tensor products of univariate basis functions).
– Extend random walks on the line to random walks on a regular grid.
f f f f f
f f f f f
f f f f f
f f f f f
f f f f f
v v
v
v
Empirical Bayes Inference in Structured Hazard Regression 12
Thomas Kneib Structured Hazard Regression
• Spatial effects for regional data s ∈ {1, . . . , S}: (Intrinsic Gaussian) Markov randomfields.
– Bivariate extension of a first order random walk on the real line.
– Define appropriate neighbourhoods for the regions.
– Assume that the expected value of fspat(s) = ξs is the average of the functionevaluations of adjacent sites:
ξs|ξs′, s′ 6= s, τ2 ∼ N
1
Ns
∑
s′∈∂s
ξs′,τ2
Ns
.
τ2
2
t−1 t t+1
f(t−1)
E[f(t)|f(t−1),f(t+1)]
f(t+1)
Empirical Bayes Inference in Structured Hazard Regression 13
Thomas Kneib Structured Hazard Regression
• Spatial effects for point-referenced data: Stationary Gaussian random fields.
– Well-known as Kriging in the geostatistics literature.
– Spatial effect follows a zero mean stationary Gaussian stochastic process.
– Correlation of two arbitrary sites is defined by an intrinsic correlation function.
– Can be interpreted as a basis function approach with radial basis functions.
Empirical Bayes Inference in Structured Hazard Regression 14
Thomas Kneib Structured Hazard Regression
• Cluster-specific frailty terms:
– Account for unobserved heterogeneity.
– Easiest case: i.i.d Gaussian frailty.
• All covariates in the discussed model terms are allowed to be piecewise constanttime-varying.
Empirical Bayes Inference in Structured Hazard Regression 15
Thomas Kneib Empirical Bayes Inference
Empirical Bayes Inference
• Generic representation of structured hazard regression models:
λ(t) = exp [x(t)′γ + f1(z1(t)) + . . . + fp(zp(t))]
• For example:
f(z(t)) = g(t) z(t) = t log-baseline effect,
f(z(t)) = u(t)g(t) z(t) = (u, t) time-varying effect of u(t),f(z(t)) = f(x(t)) z(t) = x(t) smooth function of a continuous
covariate x(t),f(z(t)) = fspat(s) z(t) = s spatial effect,
f(z(t)) = f(x1(t), x2(t)) z(t) = (x1(t), x2(t)) interaction surface,
f(z(t)) = bg z(t) = g i.i.d. frailty bg, g is a groupingindex.
• The generic representation facilitates description of inferential details.
Empirical Bayes Inference in Structured Hazard Regression 16
Thomas Kneib Empirical Bayes Inference
• All vectors of function evaluations fj can be expressed as
fj = Zjξj
with design matrix Zj, constructed from zj(t), and regression coefficients ξj.
• Generic form of the prior for ξj:
p(ξj|τ2j ) ∝ (τ2
j )−kj2 exp
(− 1
2τ2j
ξ′jKjξj
)
• Kj ≥ 0 acts as a penalty matrix, rank(Kj) = kj ≤ dj = dim(ξj).
• τ2j ≥ 0 can be interpreted as a variance or (inverse) smoothness parameter.
• Relation to penalized likelihood: Penalty terms
Pλj(ξj) = log[p(ξj|τ2
j )] = −12λjξ
′jKjξj, λj =
1τ2j
.
Empirical Bayes Inference in Structured Hazard Regression 17
Thomas Kneib Empirical Bayes Inference
• Likelihood contributions for right- and uncensored survival times:
λ(T )δ exp
(−
∫ T
0
λ(t)dt
),
where δ is the censoring indicator.
• Likelihood contributions for interval-censored observations:
P (T ∈ [Tlower, Tupper]) = S(Tlower)− S(Tupper)
= exp
[−
∫ Tlower
0
λ(t)dt
]− exp
[−
∫ Tupper
0
λ(t)dt
].
⇒ Derivatives of the log-likelihood become much more complicated for interval-censored survival times.
• In general, numerical integration has to be used to evaluate the cumulative hazardrate (e.g. the trapezoidal rule).
• Left truncation can easily be included.
Empirical Bayes Inference in Structured Hazard Regression 18
Thomas Kneib Empirical Bayes Inference
• Principal idea of empirical Bayes estimation:
– Differentiate between parameters of primary interest and hyperparameters.
– Estimate the hyperparameters up-front from their marginal posterior.
– Plug the resulting estimates back into the joint posterior and maximize with respectto the parameters of primary interest (yields posterior mode estimates).
• In structured hazard regression models:
– regression coefficients are parameters of primary interest,
– variance components are hyperparameters.
• Employ mixed model methodology to perform empirical Bayes inference: Consider ξj
a correlated random effect with multivariate Gaussian distribution.
• Problem: In most cases partially improper random effects distribution (kj = rk(Kj) <dim(ξj) = dj).
Empirical Bayes Inference in Structured Hazard Regression 19
Thomas Kneib Empirical Bayes Inference
• Mixed model representation: Decompose
ξj = Xjβj + Zjbj,
wherep(βj) ∝ const and bj ∼ N(0, τ2
j Ikj).
⇒ βj is a fixed effect and bj is an i.i.d. random effect.
• This yields the variance components model
λ(t; ·) = exp [x′β + z′b] ,
where in turnp(β) ∝ const, b ∼ N(0, Q),
andQ = blockdiag(τ2
1 I, . . . , τ2pI).
Empirical Bayes Inference in Structured Hazard Regression 20
Thomas Kneib Empirical Bayes Inference
• Obtain empirical Bayes estimates / penalized likelihood estimates via iterating
– Penalized maximum likelihood for the regression coefficients β and b.
– Restricted Maximum / Marginal likelihood for the variance parameters in Q:
Lmarg(Q) =∫
L(β, b,Q)p(b)dβdb → maxQ
.
• Penalized score function and penalized Fisher information:
sp(β, b) =
∂l(β, b)∂β
∂l(β, b)∂b
−Q−1b
Fp(β, b) =
∂2l(β, b)∂β∂β′
∂2l(β, b)∂β∂b′
∂2l(β, b)∂b∂β′
∂2l(β, b)∂b∂b′
−Q−1
.
Empirical Bayes Inference in Structured Hazard Regression 21
Thomas Kneib Empirical Bayes Inference
• Marginal likelihood estimation corresponds to REML estimation of variances inGaussian mixed models.
• The marginal likelihood can not be derived analytically ⇒ Apply a Laplaceapproximation.
• This yields the approximate marginal log-likelihood
lmarg(Q) ≈ l(β̂, b̂)− 12
log |Q| − 12b̂′Q−1b̂− 1
2log |Fp|,
where Fp is the penalised Fisher information matrix.
• If both l(β̂, b̂) and b̂ vary only slowly when changing the variance components we canfurther reduce the marginal log-likelihood to
lmarg(Q) ≈ −12
log |Q| − 12
log |Fp| − 12b′Q−1b,
where b denotes a fixed value, e.g. a current estimate.
Empirical Bayes Inference in Structured Hazard Regression 22
Thomas Kneib Empirical Bayes Inference
• This allows to device a Fisher Scoring algorithm based on matrix differentiation rules.
Empirical Bayes Inference in Structured Hazard Regression 23
Thomas Kneib Software
Software
• Implemented in BayesX, a software package for Bayesian inference in geoadditive andrelated models.
• Available from
http://www.stat.uni-muenchen.de/~bayesx
Empirical Bayes Inference in Structured Hazard Regression 24
Thomas Kneib Software
• More features:
– Fully Bayesian inference based on MCMC in comparable model classes.
– Univariate responses from exponential families (Gaussian, Binomial, Poisson,Negative Binomial, Gamma, . . . ).
– Categorical responses (multinomial regression models, cumulative models,sequential models).
• Latest development: Multi-state models.
Empirical Bayes Inference in Structured Hazard Regression 25
Thomas Kneib Multi-State Models
Multi-State Models
• Multi-state models form a general class for the description of the evolution of discretephenomena in continuous time.
• We observe paths of a process
X = {X(t), t ≥ 0} with X(t) ∈ {1, . . . , q}.
• Yields a similar data structure as for Markov processes.
• Examples:
– Recurrent events:
����3
����
����
1 2........................................................................................................................................................
.......................
........................................................................................................................................................
.......................
........................................................................................................................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
........................................................................................................................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
........................................................................................................................
.......................
........................................................................................................................
.......................
Empirical Bayes Inference in Structured Hazard Regression 26
Thomas Kneib Multi-State Models
– Disease progression:
����
q
����
����
����
����
1 2 3 q-1· · ·
.............................................................................................................
.......................
.............................................................................................................
.......................
.............................................................................................................
.......................
.............................................................................................................
.......................
.............................................................................................................
.......................
.............................................................................................................
.......................
.............................................................................................................
.......................
.............................................................................................................
.......................
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..............................................................................................................................................................................
.......................
................................................................................................................................................................................................................................................................................................
.......
.......................
.......................................................................................................................................................................................................................................................................................................
.......................
– Competing risks:
����
����
����
����
2 3 4 q
����1
· · ·
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
..................................................................................................................................................................................
.......................
...................................................................................................................................................................................................................................................................................................
.......................
.........................................................................................................................................................................................................................................................................................
.......................
• (Homogenous) Markov processes can be compactly described in terms of the transitionintensities
λij = lim∆t→0
P (X(t + ∆t) = j|X(t) = i)∆t
Empirical Bayes Inference in Structured Hazard Regression 27
Thomas Kneib Multi-State Models
• Often not flexible enough in practice since
– The transition intensities might vary over time.
– The transition intensities might be related to covariates.
– The Markov model implies independent and exponentially distributed waiting times.
Empirical Bayes Inference in Structured Hazard Regression 28
Thomas Kneib Human Sleep Data
Human Sleep Data
• Human sleep can be considered an example of a recurrent event type multi-statemodel.
• State Space:
Awake Phases of wakefulnessREM Rapid eye movement phase (dream phase)
Non-REM Non-REM phases (may be further differentiated)
• Aims of sleep research:
– Describe the dynamics underlying the human sleep process.
– Analyse associations between the sleep process and nocturnal hormonal secretion.
– (Compare the sleep process of healthy and diseased persons.)
Empirical Bayes Inference in Structured Hazard Regression 29
Thomas Kneib Human Sleep Data
stat
e
awak
eN
on−
RE
MR
EM
time since sleep onset (in hours)
cort
isol
0 2 4 6 8
050
100
150
Empirical Bayes Inference in Structured Hazard Regression 30
Thomas Kneib Human Sleep Data
• Data generation:
– Sleep recording based on electroencephalographic (EEG) measures every 30 seconds(afterwards classified into the three sleep stages).
– Measurement of hormonal secretion based on blood samples taken every 10minutes.
– A training night familiarises the participants of the study with the experimentalenvironment.
⇒ Sleep processes of 70 participants.
• Simple parametric approaches are not appropriate in this application due to
– Changing dynamics of human sleep over night.
– The time-varying influence of the hormonal concentration on the transitionintensities.
– Unobserved heterogeneity.
⇒ Model transition intensities nonparametrically.
Empirical Bayes Inference in Structured Hazard Regression 31
Thomas Kneib Specification of Transition Intensities
Specification of Transition Intensities
• To reduce complexity, we consider a simplified transition space:
Awake
Non-REM REM
Sleep
?
6
-
¾
λRN(t)
λNR(t)
λAS(t) λSA(t)
Empirical Bayes Inference in Structured Hazard Regression 32
Thomas Kneib Specification of Transition Intensities
• Model specification:
λAS,i(t) = exp[γ
(AS)0 (t) + b
(AS)i
]
λSA,i(t) = exp[γ
(SA)0 (t) + b
(SA)i
]
λNR,i(t) = exp[γ
(NR)0 (t) + ci(t)γ
(NR)1 (t) + b
(NR)i
]
λRN,i(t) = exp[γ
(RN)0 (t) + ci(t)γ
(RN)1 (t) + b
(RN)i
]
where
ci(t) =
{1 cortisol > 60 n mol/l at time t
0 cortisol ≤ 60 n mol/l at time t,
b(j)i ∼ N(0, τ2
j ) = transition- and individual-specific frailty terms.
Empirical Bayes Inference in Structured Hazard Regression 33
Thomas Kneib Counting Process Representation
Counting Process Representation
• A multi-state model with k different types of transitions can be equivalently expressedin terms of k counting processes Nh(t), h = 1, . . . , k counting these transitions.
awak
eN
on−
RE
MR
EM
700 720 740 760 780 800
awake => sleep
700 720 740 760 780 800
2122
2324
Non−REM => REM
700 720 740 760 780 800
67
89
sleep => awake
700 720 740 760 780 800
2223
2425
REM => Non−REM
700 720 740 760 780 800
45
67
Empirical Bayes Inference in Structured Hazard Regression 34
Thomas Kneib Counting Process Representation
• From the counting process representation we can derive the likelihood contributionsfor individual i:
li =k∑
h=1
[∫ Ti
0
log(λhi(t))dNhi(t)−∫ Ti
0
λhi(t)Yhi(t)dt
]
=ni∑
j=1
k∑
h=1
[δhi(tij) log(λhi(tij))− Yhi(tij)
∫ tij
ti,j−1
λhi(t)dt
].
k number of possible transitions.
Nhi(t) counting process for type h event and individual i.
Yhi(t) at risk indicator for type h event and individual i.
tij event times of individual i.
ni number of events for individual i.
δhi(tij) transition indicator for type h transition.
Empirical Bayes Inference in Structured Hazard Regression 35
Thomas Kneib Counting Process Representation
• Baseline effects:
0 2 4 6 8
−2.
5−
2.0
−1.
5−
1.0
awake −> sleep (mixed model)
0 2 4 6 8
−4.
5−
4.0
−3.
5−
3.0
−2.
5−
2.0
−1.
5
sleep −> awake (mixed model)
0 2 4 6 8
−14
−12
−10
−8
−6
−4
−2
Non−REM −> REM (mixed model)
0 2 4 6 8
−4.
5−
4.0
−3.
5−
3.0
−2.
5
REM −> Non−REM (mixed model)
Empirical Bayes Inference in Structured Hazard Regression 36
Thomas Kneib Counting Process Representation
• Time-varying effects for a high level of cortisol:
0 2 4 6 8
−4
−3
−2
−1
01
2
Non−REM −> REM (mixed model)
0 2 4 6 8
−2
−1
01
REM −> Non−REM (mixed model)
• Individual-specific variation is only detected for the transition between REM andNon-REM.
Empirical Bayes Inference in Structured Hazard Regression 37
Thomas Kneib Conclusions
Conclusions
• Unified framework for general regression models describing the hazard rate of survivalmodels.
• Empirical Bayes inference based on mixed model methodology.
• Extendable to models for transition intensities in multi state models.
• Future work:
– More general censoring mechanisms for multi-state models.
– Conditions for propriety of posteriors.
– Joint modelling of covariates and duration times.
Empirical Bayes Inference in Structured Hazard Regression 38
Thomas Kneib References
References
• Brezger, Kneib & Lang (2005). BayesX: Analyzing Bayesian structured additiveregression models. Journal of Statistical Software, 14 (11).
• Fahrmeir, Kneib & Lang (2004). Penalized structured additive regression forspace-time data: a Bayesian perspective. Statistica Sinica 14, 731–761.
• Kneib & Fahrmeir (2006). A mixed model approach for geoadditive hazardregression. Scandinavian Journal of Statistics, to appear.
• Kneib (2006). Geoadditive hazard regression for interval censored survival times.Computational Statistics and Data Analysis, 51, 777-792.
• Kneib & Hennerfeind (2006). Bayesian semiparametric multi-state models. Inpreparation.
Empirical Bayes Inference in Structured Hazard Regression 39