Survival Analysis I (CHL5209H) Olli Saarela Likelihood construction under non- informative censoring Piecewise constant hazard model 22-1 Survival Analysis I (CHL5209H) Olli Saarela Dalla Lana School of Public Health University of Toronto [email protected]January 14, 2015
22
Embed
Survival Analysis I (CHL5209H) · Survival Analysis I (CHL5209H) Olli Saarela Likelihood construction under non-informative censoring Piecewise constant hazard model 22-3 Parametrized
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
I The observed data consists of realizations of randomvectors (Ti ,Ei ,Zi ), i = 1, . . . , n.
I The likelihood contribution of a given individual i ,conditional on the covariates, is the probability of an eventof type ei occurring at time ti , that is,
I The individual level contributions are assumedconditionally independent, giving the likelihood function
L(θ)
≡n∏
i=1
1∏j=0
λij(ti ; θ)1{ei=j} dt exp
−∫ ti
0
1∑j=0
λij(u; θ)du
.
I Note that this is a function of both the hazard functionλi1(t; θ) characterizing the events of interest, and λi0(t; θ)characterizing the censoring events.
I We don’t want to estimate the latter; how can weeliminate it from the likelihood expression?
I We need to make further assumptions concerning thecensoring mechanism.
I Non-informative censoring is a rather abstract property, soexamples will follow.
I Note that the independence of the censoring mechanismwas conditional on the covariates zi .
I The non-informative censoring assumption is satisfied ifwe can condition on all common determinants of thecensoring events and the events of interest.
I Example: suppose that incident myocardial infarction (MI)events censor the follow-up for incident ischemic stroke(IS) events. What are the common determinants of MIand IS?
I For this, we need to parametrize the hazard functionthrough a regression equation.
I For example,
λi1(u; θ1) = exp{α + β′Zi},
where θ1 = (α, β) would specify a Poisson regressionmodel, with the baseline hazard given by exp{α} and theregression coefficients having interpretation as log-rateratios (this is a special case of a proportional hazardsmodel).
I Usually, we would not want to assume the hazard to beconstant over time.
I A generalization of this model is obtained if we assumethat hazard to be constant over pre-specified intervals.
I This also allows us to easily incorporate more than onetime scale.
I The Lexis diagram depicted the follow-up for totalmortality of 9029 individuals recruited as a cross-sectionalcohort in 1982 (then of age 25-65) until the end of year2010.
I Assume that the mortality rate is constant within theagegroups k = 1, . . . , 9 in the Lexis diagram, and withinone-year calendar time intervals l = 1, . . . , 29.
I Let dijkl ∈ {0, 1} denote whether individual i experienceda death due to cause j at age k in year j .
I Let yikl denote the person-years individual i contributed inage group k and year l .
I If we have no other individual level information, the hazardrate of any individual i in age group k and year l isassumed to be λjkl .
I This is why the model is called piecewise constant.
I We would get the same likelihood expression if we assumethe total number of deaths of type j in each agegroup/year to be independently Poisson distributed as
n∑i=1
dijkl ∼ Poisson
(λjkl
n∑i=1
yikl
).
I Thus, the model can be fitted using any available Poissonregression software (as examples, we consider the glm
I We can allow the piecewise constant hazard rates tofurther depend on individual-level covariates Zi , in whichcase the likelihood expression is of the form
n∏i=1
9∏k=1
29∏l=1
[λdijklijkl exp {−λijklyikl}
].
I While there are no Poisson distributed counts here, themodel can still be fitted as a Poisson regression. (Why?)
I This model involves only 9 + 29 + 2 = 40 parameters.
I Now we are mainly interested in the calendar time effectparameters βjl , l = 1, . . . , 29.
I Adjustment for age through the parameters αjk ,k = 1, . . . , 9 is needed to exctract the calendar time effect.(What would happen to the calendar time effect if we didnot adjust for age?)
I Note that for this model to be identifiable, one parameterrestriction, for example
I Estimating 29 calendar time effect parameters separatelyis still quite many, as there are not a large number ofdeaths during any given year in this cohort.
I Moreover, the estimated effect will not be smooth.
I If we are Bayesians (in this course, we mostly won’t be),we can force these parameters to be dependent through aprior distribution.
I The motivation for this would be to obtain more smoothestimates by borrowing strength from the previous years.