Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Gerrit Toenges

Gerrit Toenges

SAM Conference 2017

July 04, 2017
Tweet

More Decks by SAM Conference 2017

Other Decks in Research

Transcript

  1. Recurrent event analysis in the presence of an associated terminal

    event: Bias of marginal model estimates Gerrit Toenges Institute of Medical Biostatistics, Epidemiology and Informatics (IMBEI) University Medical Center Mainz July 4th 2017
  2. Motivation: Val-HeFT-trial Valsartan Heart Failure Trial1 (Val-HeFT) Randomised clinical trial

    on the efficacy of Valsartan in chronic heart failure (HF): Treatment group (n = 2511): Standard therapy + Valsartan Control group (n = 2499) : Standard therapy + Placebo Endpoints: 1. Mortality (time to death) 2. Morbidity (time to recurrent hospitalisations) Pat. 1 Pat. 2 Randomisation End of study = Hospitalisation = Death = Censoring (administrative) 1 Cohn, J.R. et al. (2001): A Randomized Trial of the Angiotensin-Receptor Blocker Valsartan in Chronic Heart Failure. N Engl J Med; 345:1667-1675 2
  3. Modelling approaches λ1 (t) = Hospitalisation rate of survivors λ2

    (t) = Mortality rate of survivors → Poisson-assumption: Both rates independent of event history 0 H1 H2 H3 ... D D D ... λ1 (t) λ2 (t) λ1 (t) λ2 (t) λ1 (t) λ2 (t) λ1 (t) λ2 (t) time t H = Hosp. D = Death Marginal Models (Conditional independent processes) λ1 (t|x) = λ10 (t) exp(βT 1 x) λ2 (t|x) = λ20 (t) exp(βT 2 x) Joint Frailty Model (Associated processes) λ1 (t|Z, x) = Zλ10 (t) exp(βT 1 x) λ2 (t|Z, x) = Zλ20 (t) exp(βT 2 x) 3
  4. Val-HeFT-trial results Analysis of Val-HeFT-data: Marginal Models vs. Joint Frailty

    Model 0.6 0.8 1.0 1.2 1.4 unpublished result: not shown unpublished result: not shown Hazard Ratio (recurrent) AG−Model JF−Model Hazard Ratio (death) Cox−Model JF−Model Hazard Ratio Results from Cohn, J.N. (2001): A Randomized Trial of the Angiotensin-Receptor Blocker Valsartan in Chronic Heart Failure. NEJM, No. 345, p. 1667-1675, additional unpublished results from Dr. M. Akacha (Novartis Pharma) 4
  5. Question: Bias of misspecified marginal model estimates True data-generating model:

    Joint Frailty Model Hospitalisations : λ1 (t|Z, x) = Zλ10 (t) exp(βT 1 x) Mortality : λ2 (t|Z, x) = Zλ20 (t) exp(βT 2 x) Analysis by misspecified marginal models Hospitalisations : λ1 (t|x) = λ10 (t) exp(βT 1 x) → Andersen-Gill model Mortality : λ2 (t|x) = λ20 (t) exp(βT 2 x) → Cox model Maximum partial likelihood estimation: βAG 1 = argmax β∈Rp n i=1 j:Tij <min{Ci ,Di } exp(βT Xi ) k∈R(Tij ) exp(βT Xk ) 5
  6. Marginal properties of the Joint Frailty Model Non-proportional marginal hospitalisation-hazards

    Gamma-Frailty, exp(β2) = 0.61, exp(β1) = 0.74 λ10(t) = const1, λ20(t) = const2 0.0 0.5 1.0 1.5 2.0 2.0 2.5 3.0 a Marginaler Hazard Control (x=0) Treatment (x=1) 0.0 0.5 1.0 1.5 2.0 a 0.74 0.76 0.78 0.80 time Marginal Hazard Ratio Marginal Hazard Marginal Hazard λ1(t|x) = λ10(t) exp(β1x) E(Z|x, D > t) decreasing in t Marginal Hazard Ratio λ1(t|x = 1) λ1(t|x = 0) = exp(β1) E(Z|x = 1, D > t) E(Z|x = 0, D > t) 6
  7. Main results The Least False Parameter βAG 1 β∗ 1

    (Andersen-Gill-estimator) (Least False Parameter) PJoint Frailty 1. 2. The Least False Parameter β∗ 1 is the root of the function f (c) = ∞ 0 ¯ y(0)(t)¯ y(1)(t) ¯ y(0)(t) + ¯ y(1)(t) exp(c) λ1 (t|x = 1) − exp(c)λ1 (t|x = 0) dt where ¯ y(i)(t) = P(X = i, D > t, C > t). Asymptotic bias of βAG 1 depends on: - Frailty-distribution - Censoring distribution - Treatment-effects - Baseline-hazards 7
  8. Bias of the Andersen-Gill-estimator Scenario: Gamma-Frailty with variance θ, λ10(t)

    = const1, λ20(t) = const2 −1.0 −0.5 0.0 0.5 1.0 β2 (Treatment−effect on mortality) Bias β 1 ∗ − β1 −0.15 0.00 0.15 θ = 0 8
  9. Bias of the Andersen-Gill-estimator Scenario: Gamma-Frailty with variance θ, λ10(t)

    = const1, λ20(t) = const2 −1.0 −0.5 0.0 0.5 1.0 β2 (Treatment−effect on mortality) Bias β 1 ∗ − β1 −0.15 0.00 0.15 θ = 0 θ = 0.5, FU 2 years 9
  10. Bias of the Andersen-Gill-estimator Scenario: Gamma-Frailty with variance θ, λ10(t)

    = const1, λ20(t) = const2 −1.0 −0.5 0.0 0.5 1.0 β2 (Treatment−effect on mortality) Bias β 1 ∗ − β1 −0.15 0.00 0.15 θ = 0 θ = 0.5, FU 2 years θ = 1.0, FU 2 years 10
  11. Bias of the Andersen-Gill-estimator Scenario: Gamma-Frailty with variance θ, λ10(t)

    = const1, λ20(t) = const2 −1.0 −0.5 0.0 0.5 1.0 β2 (Treatment−effect on mortality) Bias β 1 ∗ − β1 −0.15 0.00 0.15 θ = 0 θ = 0.5, FU 2 years θ = 1.0, FU 2 years θ = 0.5, FU 3 years θ = 1.0, FU 3 years 11
  12. Val-HeFT-trial results Analysis of Val-HeFT-data: Marginal Models vs. Joint Frailty

    Model 0.6 0.8 1.0 1.2 1.4 unpublished result: not shown unpublished result: not shown Hazard Ratio (recurrent) AG−Model JF−Model Hazard Ratio (death) Cox−Model JF−Model Hazard Ratio Results from Cohn, J.N. (2001): A Randomized Trial of the Angiotensin-Receptor Blocker Valsartan in Chronic Heart Failure. NEJM, No. 345, p. 1667-1675, additional unpublished results from Dr. M. Akacha (Novartis Pharma) 12
  13. Summary Direction of treatment-effect on mortality (β2 ) Bias-Direction of

    marginal recurrent effect estimator (βAG 1 ) β2 < 0 β∗ 1 > β1 (Bias positive) β2 = 0 β∗ 1 = β1 (No bias) β2 > 0 β∗ 1 < β1 (Bias negative) = Analytically shown for Gamma-Frailty = Analytically shown for every Frailty-distribution 13