Joint Modelling of Accelerated Failure Time and Longitudinal Data By Yi-Kuan Tseng Joint Work With Professor Jane-Ling Wang Professor Fushing Hsieh Tseng Y.K., Hsieh F., and Wang J.L. (2005). 92, pp. 587-603, Biometrika. CD4 count plot of five patients 60 ID 69 ID 58 ID 62 ID 64 ID 74 50 CD4 count 40 30 20 10 0 0 500 1000 1500 2000 Days 2500 3000 3500 I. Introduction W (t ) X (t ) e(t ) X (t ) : longitudinal covariates e(t ) : independent measurement error {t | X (t )} 0 (t ) exp( X (t )} X (t ) { X ( s ) : 0 s t}; : regression parameter 0 : unspecified baseline hazard rate function CD4 counts and time to AIDS (or death) X (t ) b0 b1t Self and Powitan(1992), Degruttola and Tu(1994), Tsiatis et al.(1995), Faucett and Thomas (1996), Wulfsohn and Tsiatis(1997) Bycott and Taylor(1998) Dafni and Tsiatis (1998), Tsiatis and Davidian (2001) X (t ) f (t )T b U (t ) f(t) :a vector of known functions of time t, U (t) : a stochastic process Taylor et al.(1994), Lavalley and Degruttola(1996), Henderson et al.(2000),Wang and Taylor(2001), Xu and Zeger(2001) Two-stage partial likelihood approaches -truncation causes bias Joint likelihood approaches -robust to the distribution of random effects - unbiased - efficient Bayesian approaches Conditional score approaches Accelerated failure time model is an attractive alternative when the proportional hazard assumption fails. For time independent covariates X: log T ' X e T : survival time; X : time independent covariates; e: random error Suppose S0 : baseline survival function ( T | X 0) ( ' X ) S ( t ) S { te } =S0 (u ) 0 U ~ S0 Te ' X U U : a subject would have lived if there's no exposure (X 0) u t exp( ' X ) t 2 t 30 (years old) with the same survival probability as u 60 (year old), (aging twice faster) For time dependent covariates X(t), we consider the AFT model in Cox and Oakes (1984): T U ~ S0 , where U { X (T ); } exp{ X ( s)}ds S{t | X (t )} S0[ {X (t; )}] Biological 0 meaning: Allows the influence of entire covariate history on subject specific risk. For an absolutely continuous S0, the hazard rate function with covariate history: t {t | X (t )} 0 [ e ' X ( s ) ds]e ' X (t ) 0 [ { X (t ); }] '{ X (t ); } 0 If baseline hazard is unspecified, the expression corresponds to a semi-parametric model. Robins and Tsiatis (1992)– rank estimating equation Lin and Ying (1995)– asym. consistency and Normality Hsieh (2003)– over-identified estimating equation Goal of the study: provide an effective estimators for β with unspecified baseline hazard and the parameters of longitudinal process Different assumptions on baseline hazard: -- Wulfsohn and Tsiatis (1997) Discrete baseline hazard with jumps at event times -- Our assumption: The baseline hazard is a step function. II. Joint AFT and Longitudinal model Notations: Ti : event time of subject i, i 1, ,n Ci : censoring time Vi : observed time min (Ti , Ci ) i : 1(Ti Ci ), event time indicator ti : measurement schedule (tij : tij Vi ), j 1,...mi Wi : response (Wij : tij Vi ) X i () : time dependent covariate ei : measurement error Observed data for each i: (Vi , i , Wi , t i ), independent across i. Model for longitudinal data: Wi X i (t i ) ei X i (t ) bi T (t ) (linear mixed effect model) (t ) {1 (t ),..., p (t )}T : vector of known functions of time t bi T (bi1 ,..., bip ) : p-dimensional random effects ~ N p ( , ) ei ei ~ N (0, e2 I ) Examples : p 2, {0 (t ), 1 (t )} (1, t ) p k , {0 (t ),..., p 1 (t )} (1, t ,..., t k 1 ) p 2, {0 (t ), 1 (t )} {log(t ), t 1} Model for survival: (t | X (t )) (t | , bi ) 0 ( (t; , bi ) ' (t; , bi ) Where t (t ; , bi ) e X (s) ds e 0 (t ; , bi ) e ' Joint t biT ( s ) ds, 0 X (t ) e biT ( t ) likelihood: Assumptions -- noninformative censoring -- noninformative measurement schedule tij , both are independent of future covariate history and random effects bi L( ) L( , , , e2 , 0 ) i 1[ { j 1 f (Wij | bi , t i , e2 )} f (Vi , i | bi , t i , 0 , ) f (bi | , ) dbi ] n mi f (Wij | bi , t i , e2 ) ~ N{biT ( s), e2 } f (bi | , ) ~ N ( , ) f (Vi , i | bi , t i , 0 , ) [0 ( (Vi ; , bi ) (Vi ; , bi )] exp{ ' i (Vi ; ,bi ) 0 0 (t )dt} III. EM Algorithm Complete data likelihood: L ( ) i 1[ j 1 f (Wij | bi , ti , e2 )} f (Vi , i | bi , ti , 0 , ) f (bi | , )] * n mi M-step: Let E{ h(bi ) | Vi , i ,Wi , ti , } Ei { h(bi ) } 2 e be the conditional expectation based on the current estimate ( , , , , 0 ). Dfferentiating Ei {log L* ( )} n Ei (bi ) / n, i 1 n Ei (bi )(bi )T / n, i 1 n mi n Ei {Wij biT (tij )}2 / mi 2 e i 1 j 1 i 1 For 0 : Let T1 ,..., Td denote d distinct uncensored event time The corresponding baseline survival time are: Tk uk exp{ bkT ( s)}ds, k 1,..., d 0 Estimate uk by current estimate of and the current empirical Bayes estimate of bi u ( k ) denote these estimates in ascending order--- 0 =u (0) u (1) u(d ) Therefore, we have d 0 (u ) Ck 1{u k 1 Ck For n i 1 n i 1 ( j 1) u u ( j ) } Ei [ i 1{u ( k 1) u u ( k ) } ] i Ei [{u ( k ) u ( k 1) }1{u ( k 1) u u ( k ) } ] i : Plug 0 in Ei {log L* ( )}, d d T Ei i log[ C j 1{u( j1) u u( j ) } ] i {bi (Vi )} C j {u ( k ) u ( k 1) }1{u( k 1) u u( k ) } i i 1 j 1 j 1 n n n mi i 1 i 1 j 1 Ei {log f (bi | , )} Ei { log f (Wij |bi , e2 )} no closed form expression for . We may maximize the conditional likelihood by numerical method. There’s E-step: To compute Ei (.),we need knowledge of f (bi | Vi , i ,Wi , ti , ) which can be expressed as: f (Vi , i | bi , ti , ) f (bi | Wi , ti , ) f (Vi , i | bi , ti , ) f (bi | Wi , ti , )dbi Let * { T (ti1 ) ,..., T (timi ) }T , A { (ti1 ),..., (timi )}T . * 11 12 Wi Then ~ N , , and therefore 21 22 bi 1 1 bi | Wi , ti , ~ N { 2111 (Wi A ), 22 2111 12 } To derive Ei (.), we may generating M multivariate normal sequences for bi | Wi , ti , , denoted by Ni ( Ni1 ,...NiM ) E {h(b )} i The i M j 1 h( N ij ) f (Vi , i | N ij , ti , ) M j 1 f (Vi , i | N ij , ti , ) , M is large. T accuracy increases as M increases. In order to have h higher accuracy and less computing time, we may follow the suggestion in Wei and Tanner (1990) . That is, to use small value of M in the initial iterations of the algorithm, and increase the values of M as the algorithm moves closer to convergence. We encounter two difficulties when estimating standard error of : EM algorithm involved missing information -Remedies in Louis (1982) and McLachlan and Krishnan (1997) are valid for finite dimensional parameter space. No explicit profile likelihood - Need projection onto all other parameters - However, it’s very hard to derive due to λ0 Bootstrap technique in Efron(1994): 1. Generating bootstrap sample 0* from original observed data 0 . * 2. The EM algorithm is applied to the bootstrap sample to derive the MLE . * 0 3. Repeat step 1 and 2 B times. B B 4. Compute Cov( ) 1/( B 1) ( b ) b ) , where b = / B. * b 1 * b * b T b 1 * b IV. Simulation Studies Sample size n=100 with 100 MC replications -- preliminary scheduled measurement times: (0, 1, ... , 7) -- (t ) (1, t ) -- (1,0.5)T -- 0 1, 1, e2 0.25 (i) No censoring with ( 11 , 12 , 22 ) (0.01, 0.001,0.01) (ii) With censoring time ~ exponential distribution with mean 25. (iii) With same setting except 22 0.3 and 35% negative values of bi are truncated (i) Normal random effects without censoring β μ1 μ2 σ11 σ12 σ22 σe2 target 1 1 0.5 0.01 -0.001 0.001 0.25 mean 1.0075 0.9955 0.5013 0.0087 -0.0011 0.0009 0.2528 SD 0.0945 0.0163 0.0055 0.0015 0.0002 0.0002 0.0135 (ii) Normal random effects with censoring β μ1 μ2 σ11 σ12 σ22 σe2 target 1 1 0.5 0.01 -0.001 0.001 0.25 mean 0.9918 0.9944 0.5015 0.0083 -0.0011 0.0009 0.2516 SD 0.1272 0.0249 0.0056 0.0023 0.0004 0.0002 0.0198 (iii) Nonnormal random effects with censoring β μ1 μ2 σ11 σ12 σ22 σ2e target 1 1 0.5 0.01 -0.001 0.001 0.25 empirical target 1 0.9993 0.6758 0.0104 -0.0058 0.1358 0.2753 mean 0.9950 1.0007 0.6682 0.0099 -0.0006 0.1627 0.2500 SD 0.1091 0.0140 0.0535 0.0004 0.0036 0.0318 0.0223 V. Application on Medfly data The medfly (Mediterranean fruit fly) data: --From Carey, et al. (1998) -- We focus on 251 female medflies which have the most egg reproduction (>1150). --Range of event time from 22 to 99 -- Range of total reproduction from 1151 to 2349 --No censoring and missing Relationship between daily egg laying and mortality --Violate the proportionality (By scaled Schoenfeld residual test with p-value 0.00305) Profiles of daily egg laying of first three flies 5 subject1 subject2 subject3 4.5 4 log(# of daily egg laying+1) 3.5 3 2.5 2 1.5 1 0.5 0 0 20 40 60 Time 80 100 120 Initial model: W (t ) X (t ) e (t ) * * * X (t ) t exp[b1 (t )] * b0 Log transformed model: W (t ) log[W * (t ) 1] X (t ) e(t ) X (t ) b0 log(t ) b1 (t 1) The parameter estimates derived from original data and 100 bootstrap samples under the joint AFT β μ1 μ2 σ11 σ12 σ22 σe fitted values -0.4340 2.1227 -0.1442 0.3701 -0.0482 0.0068 0.8944 bootstrap mean -0.4313 2.1112 -0.1429 0.3651 -0.0483 0.0066 0.8958 bootstrap SD 0.0115 0.0375 0.0051 0.0353 0.0002 0.0005 0.0223 Fitting incomplete medfly data: --Randomly select 1-7 days as the corresponding schedule times for each individual. --Then, add the day of death as the last schedule time. Therefore, each individual may have 2-8 repeated measurements. -- The sub data set is further censored by exponential distribution with mean 500 (20% censoring rate) The parameter estimates derived from incomplete data and 100 bootstrap samples under the joint AFT β μ1 μ2 σ11 σ12 σ22 σe fitted values -0.3890 2.2011 -0.1665 0.2833 -0.0382 0.0051 0.9775 bootstrap mean -0.3526 2.1986 -0.1575 0.2862 -0.0398 0.0057 0.9712 bootstrap SD 0.0323 0.0461 0.0074 0.0351 0.0046 0.0006 0.0570 The End