D.I. Survival models and copulas

Similar documents
Chapter Three Systems of Linear Differential Equations

Lecture 6 - Testing Restrictions on the Disturbance Process (References Sections 2.7 and 2.10, Hayashi)

N H. be the number of living fish outside area H, and let C be the cumulative catch of fish. The behavior of N H

On Multicomponent System Reliability with Microshocks - Microdamages Type of Components Interaction

The Fundamental Theorems of Calculus

23.2. Representing Periodic Functions by Fourier Series. Introduction. Prerequisites. Learning Outcomes

An introduction to evolution PDEs November 16, 2018 CHAPTER 5 - MARKOV SEMIGROUP

Lecture 2-1 Kinematics in One Dimension Displacement, Velocity and Acceleration Everything in the world is moving. Nothing stays still.

(1) x (2) x x x x. t ( ) t ( ) (2) t ( ) (1) P v p dt v p dt v p dt t 0.096t 0.096t. P e dt e dt e dt P

The consumption-based determinants of the term structure of discount rates: Corrigendum. Christian Gollier 1 Toulouse School of Economics March 2012

Seminar 5 Sustainability

Characteristics of Linear System

Online Learning with Partial Feedback. 1 Online Mirror Descent with Estimated Gradient

( ) ( ) if t = t. It must satisfy the identity. So, bulkiness of the unit impulse (hyper)function is equal to 1. The defining characteristic is

R t. C t P t. + u t. C t = αp t + βr t + v t. + β + w t

Kriging Models Predicting Atrazine Concentrations in Surface Water Draining Agricultural Watersheds

COMPETITIVE GROWTH MODEL

Technical Appendix to Modeling Movie Lifecycles and Market Share. All our models were estimated using Markov Chain Monte Carlo simulation (MCMC).

We just finished the Erdős-Stone Theorem, and ex(n, F ) (1 1/(χ(F ) 1)) ( n

Chapter 2. Models, Censoring, and Likelihood for Failure-Time Data

Diebold, Chapter 7. Francis X. Diebold, Elements of Forecasting, 4th Edition (Mason, Ohio: Cengage Learning, 2006). Chapter 7. Characterizing Cycles

Inventory Analysis and Management. Multi-Period Stochastic Models: Optimality of (s, S) Policy for K-Convex Objective Functions

Homework 2 Solutions

Unit Root Time Series. Univariate random walk

TEACHER NOTES MATH NSPIRED

Chapter 2. First Order Scalar Equations

23.5. Half-Range Series. Introduction. Prerequisites. Learning Outcomes

A Second-row Parking Paradox

Math 2142 Exam 1 Review Problems. x 2 + f (0) 3! for the 3rd Taylor polynomial at x = 0. To calculate the various quantities:

EXERCISES FOR SECTION 1.5

Integral representations and new generating functions of Chebyshev polynomials

Two-stage Benchmarking of Time-Series Models for. Small Area Estimation. Danny Pfeffermann, Richard Tiller

EE100 Lab 3 Experiment Guide: RC Circuits

Vehicle Arrival Models : Headway

( ) ( ) ( ) ( u) ( u) = are shown in Figure =, it is reasonable to speculate that. = cos u ) and the inside function ( ( t) du

CHAPTER 2 Signals And Spectra

Echocardiography Project and Finite Fourier Series

Linear Response Theory: The connection between QFT and experiments

Guest Lectures for Dr. MacFarlane s EE3350 Part Deux

Math 334 Fall 2011 Homework 11 Solutions

VS203B Lecture Notes Spring, 2011 Topic: Diffraction

Model Reduction for Dynamical Systems Lecture 6

Lecture Notes 2. The Hilbert Space Approach to Time Series

Logic in computer science

2. Nonlinear Conservation Law Equations

Information and estimation in Fokker-Planck channels

4 Sequences of measurable functions

Section 5: Chain Rule

5. Stochastic processes (1)

10. State Space Methods

Class Meeting # 10: Introduction to the Wave Equation

3.1.3 INTRODUCTION TO DYNAMIC OPTIMIZATION: DISCRETE TIME PROBLEMS. A. The Hamiltonian and First-Order Conditions in a Finite Time Horizon

15. Vector Valued Functions

Wavelet Methods for Time Series Analysis. What is a Wavelet? Part I: Introduction to Wavelets and Wavelet Transforms. sines & cosines are big waves

On Customized Goods, Standard Goods, and Competition

8. Basic RL and RC Circuits

) were both constant and we brought them from under the integral.

Impact of International Information Technology Transfer on National Productivity. Online Supplement

Nature Neuroscience: doi: /nn Supplementary Figure 1. Spike-count autocorrelations in time.

EECE251. Circuit Analysis I. Set 4: Capacitors, Inductors, and First-Order Linear Circuits

Overview. COMP14112: Artificial Intelligence Fundamentals. Lecture 0 Very Brief Overview. Structure of this course

arxiv: v1 [math.pr] 27 Dec 2018

CHERNOFF DISTANCE AND AFFINITY FOR TRUNCATED DISTRIBUTIONS *

OBJECTIVES OF TIME SERIES ANALYSIS

d 1 = c 1 b 2 - b 1 c 2 d 2 = c 1 b 3 - b 1 c 3

11!Hí MATHEMATICS : ERDŐS AND ULAM PROC. N. A. S. of decomposiion, properly speaking) conradics he possibiliy of defining a counably addiive real-valu

Convergence of the Neumann series in higher norms

Matlab and Python programming: how to get started

Matrix Versions of Some Refinements of the Arithmetic-Geometric Mean Inequality

Homework sheet Exercises done during the lecture of March 12, 2014

Simulation-Solving Dynamic Models ABE 5646 Week 2, Spring 2010

Lecture 20: Riccati Equations and Least Squares Feedback Control

When to Sell an Asset Where Its Drift Drops from a High Value to a Smaller One

on the interval (x + 1) 0! x < ", where x represents feet from the first fence post. How many square feet of fence had to be painted?

Random variables. A random variable X is a function that assigns a real number, X(ζ), to each outcome ζ in the sample space of a random experiment.

Optimality Conditions for Unconstrained Problems

STATE-SPACE MODELLING. A mass balance across the tank gives:

KINEMATICS IN ONE DIMENSION

Robust estimation based on the first- and third-moment restrictions of the power transformation model

6. Stochastic calculus with jump processes

EE 435. Lecture 35. Absolute and Relative Accuracy DAC Design. The String DAC

Basilio Bona ROBOTICA 03CFIOR 1

Physics 127b: Statistical Mechanics. Fokker-Planck Equation. Time Evolution

18 Biological models with discrete time

Section 3.5 Nonhomogeneous Equations; Method of Undetermined Coefficients

Unsteady Flow Problems

Lecture 2: Optics / C2: Quantum Information and Laser Science

Circuit Variables. AP 1.1 Use a product of ratios to convert two-thirds the speed of light from meters per second to miles per second: 1 ft 12 in

Chapter 1 Fundamental Concepts

Representation of Stochastic Process by Means of Stochastic Integrals

Solutions of Sample Problems for Third In-Class Exam Math 246, Spring 2011, Professor David Levermore

Electrical and current self-induction

Mon Apr 9 EP 7.6 Convolutions and Laplace transforms. Announcements: Warm-up Exercise:

Chapter 7 Response of First-order RL and RC Circuits

Economics 8105 Macroeconomic Theory Recitation 6

Designing Information Devices and Systems I Spring 2019 Lecture Notes Note 17

t is a basis for the solution space to this system, then the matrix having these solutions as columns, t x 1 t, x 2 t,... x n t x 2 t...

ADDITIONAL PROBLEMS (a) Find the Fourier transform of the half-cosine pulse shown in Fig. 2.40(a). Additional Problems 91

Topics in Combinatorial Optimization May 11, Lecture 22

1. Diagnostic (Misspeci cation) Tests: Testing the Assumptions

Transcription:

D- D. SURVIVAL COPULA D.I. Survival moels an copulas Definiions, relaionships wih mulivariae survival isribuion funcions an relaionships beween copulas an survival copulas. D.II. Fraily moels Use of a laen variable o inrouce epenence beween survival imes. Link wih Archimeean copula D.III. Depenence measures Paricular care shoul be pai when measuring epenence among survival imes. Properies of Kenall s au, Spearman s rho an Tail epenences in a survival seing.

D-2 D.IV. Compeing risk moels Definiion an properies D.V. Esimaion Problems of censoring an runcaion. D.VI. Conclusions

D-3 D.I. Survival moels an copulas The erm mulivariae survival aa covers he fiel where inepenence beween survival imes canno be assume. We may parallel he consrucion of mulivariae isribuion hrough he use of copulas in a survival framework. Firs we consier he univariae aa separaely in orer o characerize he specific properies of he survival imes. Then we search o escribe he join behavior of he survival imes by aking ino accoun he properies exhibie in he firs sep.

D-4 a Univariae survival noions Le T enoe a survival ime wih isribuion F an ensiy f. The survival funcion is given by [ > ] F S P T. The hazar rae or risk funcion λ is efine as λ P[ T + ΔT ] lim Δ 0 Δ. I can be inerpree as he insananeous failure rae assuming he sysem has survive o ime. I is given by f λ. S

D-5 The hazar funcion is equal o Λ λ s s. 0 I is also known uner he name: inegrae hazar funcion or cumulaive hazar funcion. We ge he relaionship : S exp Λ In some cases we can incorporae explanaory variables in he moeling of λ, an we have hen λ exp Xβ λ 0 where λ 0 is calle he baseline hazar funcion Cox proporional hazar rae moel.

D-6 b Mulivariae survival noions The previous efiniions can be exene o he mulivariae case. The mulivariae survival funcion S is efine by where,..., P T [ > T ] S,..., > T,...,T are survival imes wih S j j. univariae survival funcions We have S0,...,0,,0,...,0 S. j j j Noe ha S,..., F,...,. The ensiy is simply f,...,,..., F,..., S,...,

D-7 Mulivariae exensions of he hazar rae an he hazar funcion are given by λ max,..., lim P[ T Δ j 0 Δ... or equivalenly: + Δ Δ,... T,...] λ,..., f S,...,,..., an,...,... s,..., s s... 0 0 Λ λ s Relaionship beween S an Λ canno be simply formulae, since coniional hazar raes nee o be aken ino accoun.

D-8 Copulas are hen a naural ools o evelop mulivariae survival funcions from marginal univariae survival funcions. c Survival copulas A mulivariae survival funcion S can be represene as follows : S,..., C S,..., S, where C is a copula Sklar heorem for survival funcions. The survival copula C couples he join survival funcion o is univariae margins in a manner compleely analogous o he way a copula connecs he join isribuion funcion o is margins.

There exiss a link beween he survival C an he copula C. In he bivariae case i is given by D-9 C u, u2 u + u2 + C u, u2 Noe ha we can buil a survival funcion as S,..., C S S,..., C S,..., S,..., S or as for a given copula C. This will no yiel he same survival funcions excep in some cases. For example i can be shown ha for ellipical copulas C C normal, suen. I is also rue for he Frank copula. Then i is equivalen o work wih he copula or he survival copula.

D-0 D.II. Fraily moels The main iea is o inrouce epenence beween survival imes T,...,T by using an unobserve ranom variable W, calle he fraily. I correspons o a laen or hien variable moeling. Given he fraily W wih isribuion G he survival imes are assume o be inepenen : [ > T > W w] P T,..., j [ > W w] P T j j

D- We ake hen S,..., n w j S j W w [ ] j j j w ψ, where ψ j j is he baseline survival funcion in a proporional hazar moel: ψ j j exp Λ exp λ s s i j j 0 i The unconiional join survival funcion is furher efine as S E,..., n [ S,..., ] n W S,..., n w G w We only nee o inegrae w.r.. he isribuion G.

D-2 I can be shown ha a survival fraily copula is a special case of he consrucion base on S,..., C S,..., S where C is an Archimeean copula wih a generaor corresponing o he inverse of he Laplace ransform of he isribuion of he fraily variable. Remark ha fraily moels exhibi a PQD behavior only, which migh be an hanicap for he moeling of some aa. Recall ha an Archimeean copula is such ha C u, u2 ϕ ϕ u + ϕ u2 where ϕ is calle he generaor of he copula.

D-3 The name Archimeean comes from one of he mahemaical propery of his caegory of copula which is relae o he Archimeean axiom: if a,b are posiive real numbers, hen here exiss an ineger n such ha na>b. Examples are he Frank copula an he Gumbel copula. They fin a wie range of applicaions since hey are easy o consruc, 2 here is a large variey of copula families which belong o his class, 3 hey have nice mahemaical properies. The high egree of analyical racabiliy of he class is an avanage, bu he number of free parameers is ypically low. This migh become an hanicap in high imensions when he epenence srucure of he aa is complex.

D-4 D.III. Depenence measures a linear correlaion The raiional way of evaluaing epenence in a bivariae isribuion is by means of he sanar correlaion coefficien. This measure of epenence is naural an unproblemaic in he class of ellipical isribuions, bu i migh be misleaing in oher conexs, ypically encounere in survival aa. Here are some usual misinerpreaions of he Pearson correlaion couner-examples may be given.. T an T 2 are inepenen if an only if T, T 0. corr 2 2. corr T, T2 0 means ha here is no perfec epenence beween T an T 2.

D-5 3. for given margins, he permissible range of T, is [-,]. corr T 2 Survival aa are ypically posiive. Hence he lower boun can never been reache. I is furher ifficul o obain large range of correlaion because of he ype of isribuions generally use in survival moeling. For he Weibull, he inerval is ofen [-/3,/2] only. b Kenall s au an Spearman s rho The Kenall s au an Spearman s rho of he survival copula an is associae copula are equal.

D-6 c Tail epenence Tail epenence measures correspon o Upper ail epenence: [ > uu u] λ U lim P U 2 > u If λ U 0, ], hen upper ail epenence. If λ 0, hen no upper ail epenence. U Lower ail epenence: [ < uu u] λ L lim P U 2 < u 0 If λ L 0,], hen lower ail epenence. If λ 0, hen no lower ail epenence. L The upper ail epenence of he survival copula will give he lower ail epenence of is associae copula, an vice-versa.

D-7 Lower ail epenence in survival copula will characerize immeiae join eah, while upper ail epenence in survival copula will characerize long-erm join survival. Remark: Normal copula has no upper or lower ail epenence. Suen copula may. D.IV. Compeing risk moels Compeing risk moels correspon o he suy of any failure process in which here are ifferen causes of failures. T,...,T Le us consier survival imes. In a compeing risk moel he survival ime τ is efine by τ min T,..., T.

D-8 We have hen [ ],...,,..., min S S C T T P S τ The cf of he survival ime τ is,...,,..., F F C S S C F τ an is ensiy is given by,..., f S S C f i i i τ Explici forms can be foun for example for Weibull margins an a Gumbel copulas.

D-9

D-20

D-2 Uner an ii scheme we ge F F τ an f τ F f D.V. Esimaion The esimaion by maximum likelihoo are exacly he same as before when observaions are complee. Inee ML esimaion relies on he join ensiy of he survival imes. However ealing wih survival imes is no as simple, because recors on survival rais are ofen incomplee: survival aa are ofen censore or runcae.

D-22 Uner lef runcaion we only observe aa above a fixe hreshol. We have no informaion abou he behavior below he limi only repore losses above a given level. Uner censoring we have usually a mixure beween complee an incomplee aa. For example uner righ censoring we observe T if i is below a hreshol C or he hreshol C iself if i is above. The hreshol C may be fixe or ranom. Esimaion uner hese schemes are much more ifficul, especially when ealing wih nonparameric esimaion. For example uner lef runcaion i is impossible o ienify nonparamerically he par of he isribuion below he hreshol we have no informaion!.

D-23 D.VI. Conclusions The join behavior of survival imes can be easily moele hrough copulas. I is a powerful ool o analyze he epenence srucure among hese aa, especially because symmeric isribuions are no naural caniaes for hese aa. Esimaion proceures are also available in such a seing bu are more ifficul o implemen when censoring or runcaion mechanisms are presen.