Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Accommodating informative dropout and death: a joint modelling approach for longitudinal and semi-competing risks data Qiuju Li MRC Biostatistics Unit, Cambridge, UK qiuju.li@mrc-bsu.cam.ac.uk Joint work with Dr. Li Su Workshop on Flexible Models for Longitudinal and Survival Data with Applications in Biostatistics (Warwick, 27th-29th July, 2015) Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Outline 1 Introduction 2 Joint modelling of longitudinal and semi-competing risks data 3 Application: HERS data analysis 4 Conclusions Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Introduction longitudinal and semi-competing risks data, e.g., CD4 counts, dropout and HIV-related death in the HIV epidemiology research study (HERS). 3 4 5 6 7 8 9 10 11 25 20 15 10 12 0 1 2 3 4 5 6 7 visit (subject=100058) visit (subject=101059) death dropout & death 8 9 10 11 12 0 1 2 3 4 5 6 7 8 9 10 11 8 9 10 11 12 25 20 15 10 5 squart root of CD4 counts 12 visit (subject=100729) 0 25 5 10 15 20 Measurements dropout death 0 squart root of CD4 counts 5 squart root of CD4 counts 2 30 1 30 0 0 25 20 15 10 5 0 squart root of CD4 counts 30 dropout 30 complete data 0 1 2 3 4 5 6 7 visit (subject=100241) Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Some concepts mortal cohort; immortal cohort; (Aalen and Gunnes, 2010) longitudinal profile models: unconditional models, e.g., random-effects models f (Yi (t)); fully conditional models, e.g., f (Yi (t)|Si = s), s > t; partly conditional models, e,g., f (Yi (t)|Si > t) (Kurland and Heagerty, 2005; Kurland et al., 2009); GEE approaches; a likelihood-based joint modelling approach proposed subsequently. Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Notation scheduled repeated measurements of a longitudinal outcome Yi = (Yi 1 , . . . , YiM )′ , taken at visits 1, . . . , M, e.g., M = 12 for the HERS data; informative dropout and death dropout time denoted by Di , observed data Di∗ = min(Di , Si , Ci ), δiD = I (Di ≤ Si , Di ≤ Ci ); death time denoted by Si , observed data Si∗ = min(Si , Ci ), δiS = I (Si ≤ Ci ); Ci denotes non-informative censoring, e.g., end of study; covariates Xi , e.g., sex, treatment arm; Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Time-to-event processes Time-to-event data, time to dropout: last visit of follow-up; time to death: Ti time to death: τ 0 τ : the end of study. Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Time-to-event processes Discrete time-to-event data (mathematical attractiveness), time to dropout: last visit of follow-up; time to death: Ti time to death: 0 t1 tr −1 tr t(M−1) τ τ : the end of study. the discrete death time Si = r (Barrett et al, 2015, JRSSB). Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Joint models Joint models for the longitudinal and semi-competing risks data, T T Yij = xij β + zij bi + ǫij T T Pr (Di = r |Di ≥ r ) = 1 − Φ(xD,ir αD + γD,r WD,ir bi ) , T T Pr (Si = r |Si ≥ r ) = 1 − Φ(xS,ir αS + γS,r WS,ir bi ) Φ(·) standard normal cdf; β, αD , αS regression coefficients; γD,r , γS,r association effects; random effects bi ∼ N(0, Σb ); iid ǫij ∼ N(0, σ 2 ); WD,ir bi , WS,ir bi vectors of linear combinations of random effects, e.g., WD,ir bi = (bi 0 , bi 1 )T ; conditional independence assumption given random effects bi ; Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Likelihood function, Y Li (θ; yi , Di∗ , δiD , Si∗ , δiS ) i = YZ i ∞ f (longitudinal data|θ, bi )× −∞ Pr (dropout data|θ, bi ) × Pr (death data|θ, bi ) × f (bi |θ)dbi . Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Semi-competing risks data Four possible scenarios of the observed time-to-event data, (1) neither dropout nor death: Di∗ = d, Si∗ = s, (δiD , δiS ) = (0, 0); (2) dropout only: Di∗ = d, Si∗ = s, (δiD , δiS ) = (1, 0); (3) death only: Di∗ = d, Si∗ = s, (δiD , δiS ) = (0, 1); (4) both dropout and death: Di∗ = d, Si∗ = s, (δiD , δiS ) = (1, 1); Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions For the scenario (1), the likelihood function of observed data {yi = (yi 1 , . . . , yini )′ , Di∗ = d, δiD = 0, Si∗ = s, δiS = 0}, Li (θ; yi , Di∗ , δiD , Si∗ , δiS ) Z ∞ d Y T T = φ(yi ; xi β + zi bi , σ 2 Ini ) Φ(xD,ir αD + γD,r WD,ir bi ) −∞ s Y k=1 T T Φ(xS,ir αS + γS,r WS,ir bi )φ(bi ; 0, Σb )dbi ℓ=1 T ) =Li 1 (·\bi )Φ(d+s) (Ads + Bds hi ; 0, Id+s + Bds Hi−1 Bds closed-form likelihood (skewed normal distribution, Arnold 2009); Li 1 (·\bi ), hi , Hi , Ads , Bds function/vectors/matrices free of bi ; φ(·; µ, Σ) and Φ(d+s) (·; µ, Σ) denote multivariate normal pdf/cdf. Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Marginal mean profile conditional on being alive Unconditional population mean profile for an immortal cohort E (Yij |xij ) = xijT β; Conditional mean profile given being alive for a mortal cohort, we can compute E (Yij |xij , Si ≥ j) = xijT β + zijT E (bi |Si ≥ j). Analogously, f (bi |Si ≥ j) is a multivariate skew-normal distribution, Pr Sj > (j − 1)|bi f (bi ) , f (bi |Si ≥ j) = f bi |Si > (j − 1) = Pr (Si > (j − 1)) closed form of its expectation can be obtained. Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Statistical inference 1 Maximum likelihood-based approach (exact likelihood); R software utilising nlminb or optim. 2 Bayesian approach; implemented using WinBUGS. Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions HERS data description HIV epidemiology research study (HERS) (Smith et al., 1997) # of subjects: 850 (HIV positive at baseline) CD4 counts reviewed every 6 months up to 12 visits time-to-event data scenario (1): (δiD , δiS ) = (0, 0) scenario (2): (δiD , δiS ) = (1, 0) scenario (3): (δiD , δiS ) = (0, 1) scenario (4): (δiD , δiS ) = (1, 1) # of subjects 374 352 23 78 Objective: study the role of baseline patient characteristics (i.e., viral load, antiretroviral therapy (art), # of symptoms) on variation in longitudinal CD4 counts. Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Models proposed for the HERS data Yij = β0 + β1 visit + β2∼4 viral load + β5 symptoms + β6 art +β7∼9 visit*viral load + β10 visit ∗ symptoms + β11 visit ∗ art + bi 0 + bi 1 + ǫij Pr (D = r |D ≥ r ) = 1 − Φ(α i i D,i 0 + αD,i 1∼3 viral load + αD,i 4 symptoms +αD,i 5 art + αD,i 6 r + αD,i 7 r 2 + γD,0 bi 0 + γD,1 bi 1 ) Pr (Si = r |Si ≥ r ) = 1 − Φ(αS,i 0 + αS,i 1∼3 viral load + αS,i 4 symptoms +αS,i 5 art + αS,i 6 r + αS,i 7 r 2 + γS,0 bi 0 + γS,1 bi 1 ) Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions WinBUGS results Longitudinal process (CD4 counts) Joint modelling (WinBUGS) intercept visit viral load [1] (0-500) viral load [2] (500-5k) viral load [3] (5k-30k) symptoms art at baseline visit*viral load [1] visit*viral load [2] visit*viral load [3] visit*symptoms visit*art γD,0 γD,1 γS,0 γS,1 mean 15.11 -0.87 10.00 6.61 2.94 -0.12 -4.66 0.47 0.44 0.28 -0.05 0.12 0.03 0.44 0.13 1.23 sd 0.71 0.13 0.79 0.72 0.82 0.20 0.43 0.14 0.13 0.14 0.03 0.06 0.01 0.05 0.02 0.17 Qiuju Li 2.5% 13.76 -1.11 8.42 5.14 1.28 -0.51 -5.51 0.21 0.19 0.02 -0.11 -0.01 0.02 0.35 0.10 0.93 97.5% 16.51 -0.62 11.56 7.97 4.54 0.28 -3.83 0.74 0.69 0.55 0.01 0.23 0.04 0.54 0.16 1.58 Linear mixed effects (LME) estimate 14.59 -0.57 10.52 6.98 3.21 -0.14 -4.76 0.23 0.22 0.15 -0.03 0.16 - sd 0.70 0.12 0.79 0.74 0.81 0.21 0.43 0.13 0.12 0.13 0.03 0.06 - Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Marginal mean profiles 19 18 17 naive LME: E(Yij | xij)=xTij β joint model: E(Yij | xij)=xTij β E(Yij | xij, Si ≥ j) 16 square root of CD4 counts 20 Subjects: viral load [1], # of symptoms=1, antiretroviral therapy (art) at baseline; 0 1 2 3 4 5 Qiuju Li 6 7 8 9 10 11 12 visit Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Conclusions 1 a likelihood-based approach to capture the partly conditional mean profiles, accommodating both informative dropout and death; 2 offer inference for both mortal and immortal cohort; 3 a new model for semi-competing risks data in the joint modelling framework; 4 approach demonstrated by an analysis of the HERS data. Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Key references 1 Aalen, Odd O., and Nina Gunnes. ”A dynamic approach for reconstructing missing longitudinal data using the linear increments model.” Biostatistics 11, no. 3 (2010): 453-472. 2 Arnold, Barry C. ”Flexible univariate and multivariate models based on hidden truncation.” Journal of Statistical Planning and Inference 139, no. 11 (2009): 3741-3749. 3 Barrett, Jessica, Peter Diggle, Robin Henderson and David TaylorRobinson. ”Joint modelling of repeated measurements and timetoevent outcomes: flexible model specification and exact likelihood inference.” Journal of the Royal Statistical Society: Series B (Statistical Methodology) 77, no. 1 (2015): 131-148. 4 Kurland, Brenda F., and Patrick J. Heagerty. ”Directly parameterized regression conditioning on being alive: analysis of longitudinal data truncated by deaths.” Biostatistics 6, no. 2 (2005): 241-258. Qiuju Li Joint modelling of longitudinal and semi-competing risks data Introduction Joint modelling of longitudinal and semi-competing risks data Application: HERS data analysis Conclusions Thank You! Authors: Qiuju Li & Li Su MRC Biostatistics Unit, Cambridge, UK Qiuju Li Joint modelling of longitudinal and semi-competing risks data