ST745: Survival Analysis: Cox-PH!

Similar documents
ST495: Survival Analysis: Maximum likelihood

You know I m not goin diss you on the internet Cause my mama taught me better than that I m a survivor (What?) I m not goin give up (What?

Semiparametric Regression

ST495: Survival Analysis: Hypothesis testing and confidence intervals

ST745: Survival Analysis: Nonparametric methods

STAT331. Cox s Proportional Hazards Model

STAT 6350 Analysis of Lifetime Data. Failure-time Regression Analysis

In contrast, parametric techniques (fitting exponential or Weibull, for example) are more focussed, can handle general covariates, but require

ST745: Survival Analysis: Parametric

STAT 331. Accelerated Failure Time Models. Previously, we have focused on multiplicative intensity models, where

Survival Analysis Math 434 Fall 2011

β j = coefficient of x j in the model; β = ( β1, β2,

Modelling geoadditive survival data

MAS3301 / MAS8311 Biostatistics Part II: Survival

Multivariate Survival Analysis

Survival Analysis for Case-Cohort Studies

Survival Regression Models

Survival Analysis. Stat 526. April 13, 2018

Tied survival times; estimation of survival probabilities

Analysis of Time-to-Event Data: Chapter 4 - Parametric regression models

Statistical Inference and Methods

Chapter 4 Regression Models

Cox s proportional hazards model and Cox s partial likelihood

Accelerated Failure Time Models

Beyond GLM and likelihood

Cox s proportional hazards/regression model - model assessment

University of California, Berkeley

Lecture 22 Survival Analysis: An Introduction

Efficiency of Profile/Partial Likelihood in the Cox Model

Longitudinal + Reliability = Joint Modeling

Lecture 12. Multivariate Survival Data Statistics Survival Analysis. Presented March 8, 2016

From semi- to non-parametric inference in general time scale models

UNIVERSITY OF CALIFORNIA, SAN DIEGO

Cox regression: Estimation

Ph.D. course: Regression models. Introduction. 19 April 2012

Residuals and model diagnostics

Time-dependent coefficients

Extensions of Cox Model for Non-Proportional Hazards Purpose

Part [1.0] Measures of Classification Accuracy for the Prediction of Survival Times

Ph.D. course: Regression models. Regression models. Explanatory variables. Example 1.1: Body mass index and vitamin D status

On the Breslow estimator

Other Survival Models. (1) Non-PH models. We briefly discussed the non-proportional hazards (non-ph) model

FULL LIKELIHOOD INFERENCES IN THE COX MODEL

PENALIZED LIKELIHOOD PARAMETER ESTIMATION FOR ADDITIVE HAZARD MODELS WITH INTERVAL CENSORED DATA

Technical Report - 7/87 AN APPLICATION OF COX REGRESSION MODEL TO THE ANALYSIS OF GROUPED PULMONARY TUBERCULOSIS SURVIVAL DATA

Philosophy and Features of the mstate package

Power and Sample Size Calculations with the Additive Hazards Model

Efficient Semiparametric Estimators via Modified Profile Likelihood in Frailty & Accelerated-Failure Models

Proportional hazards regression

Outline. Frailty modelling of Multivariate Survival Data. Clustered survival data. Clustered survival data

Joint Modeling of Longitudinal Item Response Data and Survival

Chapter 17. Failure-Time Regression Analysis. William Q. Meeker and Luis A. Escobar Iowa State University and Louisiana State University

Statistics 262: Intermediate Biostatistics Non-parametric Survival Analysis

Analysis of Time-to-Event Data: Chapter 6 - Regression diagnostics

Validation. Terry M Therneau. Dec 2015

The coxvc_1-1-1 package

Multistate models and recurrent event models

Lecture 8 Stat D. Gillen

3003 Cure. F. P. Treasure

Regularization in Cox Frailty Models

Faculty of Health Sciences. Regression models. Counts, Poisson regression, Lene Theil Skovgaard. Dept. of Biostatistics

Relative-risk regression and model diagnostics. 16 November, 2015

log T = β T Z + ɛ Zi Z(u; β) } dn i (ue βzi ) = 0,

Frailty Models and Copulas: Similarities and Differences

Optimal Treatment Regimes for Survival Endpoints from a Classification Perspective. Anastasios (Butch) Tsiatis and Xiaofei Bai

Goodness-Of-Fit for Cox s Regression Model. Extensions of Cox s Regression Model. Survival Analysis Fall 2004, Copenhagen

Part IV Extensions: Competing Risks Endpoints and Non-Parametric AUC(t) Estimation


Multistate models and recurrent event models

Lecture 7 Time-dependent Covariates in Cox Regression

Lecture 3. Truncation, length-bias and prevalence sampling

GOV 2001/ 1002/ E-2001 Section 10 1 Duration II and Matching

Dynamic Prediction of Disease Progression Using Longitudinal Biomarker Data

Faculty of Health Sciences. Cox regression. Torben Martinussen. Department of Biostatistics University of Copenhagen. 20. september 2012 Slide 1/51

Support Vector Hazard Regression (SVHR) for Predicting Survival Outcomes. Donglin Zeng, Department of Biostatistics, University of North Carolina

Analysing geoadditive regression data: a mixed model approach

Meei Pyng Ng 1 and Ray Watson 1

Statistical Methods for Alzheimer s Disease Studies

Introduction to Statistical Analysis

A fast routine for fitting Cox models with time varying effects

Definitions and examples Simple estimation and testing Regression models Goodness of fit for the Cox model. Recap of Part 1. Per Kragh Andersen

Missing Covariate Data in Matched Case-Control Studies

Survival analysis in R

Logistic Regression. Advanced Methods for Data Analysis (36-402/36-608) Spring 2014

Logistic regression: Miscellaneous topics

Lecture 9. Statistics Survival Analysis. Presented February 23, Dan Gillen Department of Statistics University of California, Irvine

Problem Set 3: Bootstrap, Quantile Regression and MCMC Methods. MIT , Fall Due: Wednesday, 07 November 2007, 5:00 PM

An Analysis on the Efficiency of Non-Parametric Estimation of Duration Dependence by Hazard-Based Duration Model

Goodness-of-fit test for the Cox Proportional Hazard Model

Poisson regression 1/15

STAT331. Combining Martingales, Stochastic Integrals, and Applications to Logrank Test & Cox s Model

Extensions of Cox Model for Non-Proportional Hazards Purpose

( t) Cox regression part 2. Outline: Recapitulation. Estimation of cumulative hazards and survival probabilites. Ørnulf Borgan

Fitting Cox Regression Models

FULL LIKELIHOOD INFERENCES IN THE COX MODEL: AN EMPIRICAL LIKELIHOOD APPROACH

5. Parametric Regression Model

REGRESSION ANALYSIS FOR TIME-TO-EVENT DATA THE PROPORTIONAL HAZARDS (COX) MODEL ST520

ECON 721: Lecture Notes on Duration Analysis. Petra E. Todd

Full likelihood inferences in the Cox model: an empirical likelihood approach

Variable Selection in Competing Risks Using the L1-Penalized Cox Model

Transcription:

ST745: Survival Analysis: Cox-PH! Eric B. Laber Department of Statistics, North Carolina State University April 20, 2015

Rien n est plus dangereux qu une idee, quand on n a qu une idee. (Nothing is more dangerous than an idea, when you have only one idea.) Kayne West I would never want a book s autograph. I am a proud non-reader of books. Kayne West

Then and now Last time we discussed parametric regression models AFT models PH models IPCW least squares Today we ll discuss Cox-PH model Estimation and inference R code

Warm-up Explain to your stat buddy 1. Explain how a parametric regression model might be used with survival data 2. Give an example of a semi-parametric regression model 3. Name three common complaints about non-parametric regression models True or false: (T/F) The original Cox-PH paper is the most highly cited statistics paper appearing in the Journal of the Royal Statistical Society Series B? (T/F) D.R. Cox founded the statistics department at NCSU (T/F) By leaving uninteresting parts of the general model completely unspecified, semi-parametric models do not require model diagnostics (T/F) Merlin is older in dog years than Madeline is in hamster years?

D.R. Cox Sir David Cox (knighted 1985) Conceived of the Cox-PH model in a fevered dream 1 Most real life statistical problems have one or more nonstandard features. There are no routine statistical questions; only questionable statistical routines. D.R. Cox 1 Well, he was reportedly very ill anyway.

Cox-PH model Let T denote survival time and x R p covariates, Cox-PH model assumes h(t x) = h 0 (t) exp {x β}, where β R p are coefficients and h 0 (t) is left unspecified Estimation and inference does not rely on form of h 0 (t) Est and inf does depend on multiplicative form Model diagnostics are essential to validate this assumption

Cox-PH model cont d Some implications of the PH model Survivor fn S(t bx) { = {S 0 (t)} exp(x β), where S 0 (t) = exp } t 0 h 0(s)ds Subjects with covariates x and x have hazard ratio h(t x) h(t x ) = exp {(x x ) β}, thus, log {h(t x)/h(t x )} = (x x ) β βj is the per-unit increase in the log hazard (why?)

Estimation Recall that a partial likelihood is a LH with nuisance terms removed Cox (1975) derived a partial LH for estimation of Cox-PH Mathematical rigor and asymptotic results provided by our own Tsiatis (1981) Just when you thought you were done...recall dn i (t) = 1 Ti [t,t+ t),δ i =1 Yi (t) = 1 Ti t dn (t) = n i=1 dn i(t) dn(t) = (dn1 (t),..., dn n (t)) Let t (1),..., t (K) denote unique failure times, assume one failure per time Let x (i) denote covariates for subject with failure time t (i) Let R (i) denote the set of individuals at risk at t (i)

Estimation cont d We showed in L3 that under non-informative right-censoring L = P { dn(t) H(t) } t=0 P {dn(t) dn (t), H(t)} P {dn (t) H(t)}, t=0 where H(t) is the history of failure/censoring process on (0, t) Drop terms P {dn (t) H(t)} to obtain partial LH

Estimation cont d Since we assume unique failure times dn (t) is 0 or 1 given t sufficiently small If dn (t) = 0 then dn(t) 0 If dn (t) = 1 then P {dn i (t) = 1 dn (t) = 1, H(t)} = Y i (t)h(t x i ) t n k=1 Y k(t)h(t x k ) t Y i (t)exp {x i β} n k=1 Y k(t) exp {x k β, } thus, the partial LH is L(β) K i=1 exp {x (i) β } l R (i) exp {x l β}

Try it! Compute the partial likelihood Patient ID t δ z 1 2 1 2 2 2 0 2 3 3 1 1 4 4 1 3

Estimation and inference Estimator β n = arg max β L(β) L(β) is not the LH we re used to (why?), however, we ll see it can be used to conduct tests and inference Show that L(β) K i=1 } exp {x (i) β l R (i) exp { x l β} = ( n exp { x i β} ) δi n l=1 Y l(t i ) exp { x l β}, Note* above is defined with multiple failures at a given time i=1

Estimation and inference cont d Log partial likelihood (hereafter, log-lh) [ ( n n δ i x i β log Y l (t i ) exp { x l β})] i=1 l=1 Score fn U(β) = n δ i [x i x(t i, β)], i=1 where x(t, β) = n l=1 x ly l (t) exp { x l β} / n l=1 Y l(t) exp { x l β} coxph in R computes β n as solution to U(β) = 0

Yet even more estimation and inference Information matrix { n n l=1 I (β) = δ Y l(t i ) exp { x l β} } [x l x(t i, β)] [x l x(t i, β)] i n l=1 Y l(t i ) exp { x l β} i=1 Asymptotic inference based on I 1/2 ( β n β) N(0, I p ) We can also use LRT treating L as a likelihood Λ(β) = 2l( β) 2l(β)

Fitting Cox-PH in R coxph.r

Stratified Cox-PH Consider categorical covariate z (e.g,. sex) Adjust nonparametrically for by fitting h(t x, z) = h j (t) exp {x β}, where h j (t) is the baseline hazard for category j of z How is this different than including z as a predictor?