16
Survival Analysis 1 Treat everyone with sincerity, they will certainly appear likeable and friendly.

Treat everyone with sincerity, they will certainly appear likeable and friendly

Embed Size (px)

DESCRIPTION

Treat everyone with sincerity, they will certainly appear likeable and friendly. Survival Analysis. Parametric Regression Models. Abbreviated Outline. Proportional hazards (PH) modeling Accelerated failure time (AFT) modeling Diagnosis for models/ model selection. Notation. - PowerPoint PPT Presentation

Citation preview

Page 1: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis1

Treat everyone with sincerity, they will certainly appear likeable and friendly.

Page 2: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis2

Survival Analysis

Parametric Regression Models

Page 3: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis3

Abbreviated Outline

Proportional hazards (PH) modeling Accelerated failure time (AFT)

modeling Diagnosis for models/ model selection

Page 4: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis4

Notation

Y: survival time X: covariate vector hx(y): the hazard function of Y given X Sx(y): the survival function of Y given X Yx: Y given X

Page 5: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis5

Proportional Hazards Model

hx(y) = h0(y)*g(X)

Hazard functionof Y given X

Baseline hazard function

A positive function

Common choice of g(x):

)exp()( XXg T

Page 6: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis6

Accelerated Failure Times Model

Yx * g(X) = Y0

Sx(y) = S0(yg(X))

Baseline survival function

Common choice of g(x):

)exp()( XXg T

Page 7: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis7

Notes

AFT model = PH model if and only if the survival time is Weibull distributed.

A more robust (semi-parametric) method has been developed for the PH model and so fitting the parametric PH model will not be demonstrated here.

Page 8: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Several AFT Models

Weibull AFT model

Lognormal AFT model

Survival Analysis8

Page 9: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis9

Model Diagnosis

SAS reference: SAS textbook Chapter 4

1. Checking the parametric model for Y

2. Checking the AFT assumption

3. Residual analysis

Page 10: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis10

Model Diagnosis

Checking the model for Y: If no censored observations, use Q-Q

plots. If with censored observations,

compare to the K-M estimates.

Page 11: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis11

Graphical Diagnosis for Parametric Models on Y Exponential model Weibull model Lognormal model Log logistic model (exercise)

Note: these methods do not take covariates into account; must be done by groups

Page 12: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis12

Model Diagnosis

Checking the AFT model:

1. Fit Kaplan-Meier estimator to each group separately

2. Compute a sequence of percentiles for each group

3. Draw the Q-Q plot of one group vs. another group

4. “almost linear” implies AFT model

Page 13: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis13

Final Model Selection

Parametric model comparisons:

Use likelihood ratio test (See SAS textbook p.89 for details and examples)

Use AIC (See Klein Sec. 12.4)

Page 14: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis14

Residual Analysis

Cox-Snell residual:

and

are i.i.d. exp(1).

)|(log iii xySe

sei

Page 15: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis15

Residual Analysis

See SAS textbook p.95 for SAS code.

The residual analysis is NOT sensitive to the difference in model fit.

Page 16: Treat everyone with sincerity,  they will certainly appear likeable and friendly

Survival Analysis16

Summary

1. Fit AFT model including all covariates based on the Lognormal, Weibull and Generalized Gamma models for Y (totally 3 models)

2. Use LR tests/AIC to determine your initial model (either lognormal or weibull)

3. Do backward model selection and residual analysis