Multistate Modeling and Applications

Similar documents
Bayesian Methods for Machine Learning

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

Lecture 22 Survival Analysis: An Introduction

Statistics 262: Intermediate Biostatistics Non-parametric Survival Analysis

Multi-state Models: An Overview

Statistical Inference and Methods

Survival Analysis. Stat 526. April 13, 2018

A COMPARISON OF POISSON AND BINOMIAL EMPIRICAL LIKELIHOOD Mai Zhou and Hui Fang University of Kentucky

Survival Regression Models

Lecture 6 PREDICTING SURVIVAL UNDER THE PH MODEL

Step-Stress Models and Associated Inference

ABC methods for phase-type distributions with applications in insurance risk problems

Cox s proportional hazards model and Cox s partial likelihood

Semiparametric Regression

STATISTICAL INFERENCE IN ACCELERATED LIFE TESTING WITH GEOMETRIC PROCESS MODEL. A Thesis. Presented to the. Faculty of. San Diego State University

Survival Analysis: Weeks 2-3. Lu Tian and Richard Olshen Stanford University

Practice Exam 1. (A) (B) (C) (D) (E) You are given the following data on loss sizes:

11 Survival Analysis and Empirical Likelihood

e 4β e 4β + e β ˆβ =0.765

Statistical Analysis of Spatio-temporal Point Process Data. Peter J Diggle

Survival Analysis Math 434 Fall 2011

Philosophy and Features of the mstate package

Estimation for Modified Data

Constrained estimation for binary and survival data

Nonparametric Model Construction

Lecture 3. Truncation, length-bias and prevalence sampling

Chapter 4 Fall Notations: t 1 < t 2 < < t D, D unique death times. d j = # deaths at t j = n. Y j = # at risk /alive at t j = n

Survival Analysis I (CHL5209H)

Kernel density estimation in R

ST745: Survival Analysis: Nonparametric methods

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

A general mixed model approach for spatio-temporal regression data

Introduction to Reliability Theory (part 2)

Statistical Inference on Constant Stress Accelerated Life Tests Under Generalized Gamma Lifetime Distributions

Introduction to Statistical Analysis

Survival Analysis. Lu Tian and Richard Olshen Stanford University

Statistical Analysis of Competing Risks With Missing Causes of Failure

Frailty Models and Copulas: Similarities and Differences

Understanding product integration. A talk about teaching survival analysis.

Multistate models and recurrent event models

MAS3301 / MAS8311 Biostatistics Part II: Survival

Censoring and Truncation - Highlighting the Differences

Lecture 5 Models and methods for recurrent event data

Chapter 2 Inference on Mean Residual Life-Overview

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

Duration Analysis. Joan Llull

ST495: Survival Analysis: Maximum likelihood

Bayesian Modeling of Accelerated Life Tests with Random Effects

Analysis of Gamma and Weibull Lifetime Data under a General Censoring Scheme and in the presence of Covariates

Fast Likelihood-Free Inference via Bayesian Optimization

Logistic regression model for survival time analysis using time-varying coefficients

Prerequisite: STATS 7 or STATS 8 or AP90 or (STATS 120A and STATS 120B and STATS 120C). AP90 with a minimum score of 3

PhD course in Advanced survival analysis. One-sample tests. Properties. Idea: (ABGK, sect. V.1.1) Counting process N(t)

Multi-state models: prediction

5. Parametric Regression Model

Survival analysis in R

Multistate models and recurrent event models

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

Robust estimates of state occupancy and transition probabilities for Non-Markov multi-state models

Reliability Engineering I

Multivariate Survival Data With Censoring.

4 Testing Hypotheses. 4.1 Tests in the regression setting. 4.2 Non-parametric testing of survival between groups

TMA 4275 Lifetime Analysis June 2004 Solution

Exercises. (a) Prove that m(t) =

9. Estimating Survival Distribution for a PH Model

STAT331. Cox s Proportional Hazards Model

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?

Modelling of Dependent Credit Rating Transitions

1 Glivenko-Cantelli type theorems

Size and Shape of Confidence Regions from Extended Empirical Likelihood Tests

Application of Time-to-Event Methods in the Assessment of Safety in Clinical Trials

Inference for partially observed stochastic dynamic systems: A new algorithm, its theory and applications

Product-limit estimators of the survival function with left or right censored data

Analysis of Time-to-Event Data: Chapter 2 - Nonparametric estimation of functions of survival time

A new iterated filtering algorithm

Modelling geoadditive survival data

8. Parametric models in survival analysis General accelerated failure time models for parametric regression

Hybrid Censoring; An Introduction 2

ADVANCED STATISTICAL ANALYSIS OF EPIDEMIOLOGICAL STUDIES. Cox s regression analysis Time dependent explanatory variables

Estimating Load-Sharing Properties in a Dynamic Reliability System. Paul Kvam, Georgia Tech Edsel A. Peña, University of South Carolina

Analysis of Progressive Type-II Censoring. in the Weibull Model for Competing Risks Data. with Binomial Removals

New mixture models and algorithms in the mixtools package

CIMAT Taller de Modelos de Capture y Recaptura Known Fate Survival Analysis

Analysis of competing risks data and simulation of data following predened subdistribution hazards

Meei Pyng Ng 1 and Ray Watson 1

STAT Sample Problem: General Asymptotic Results

Lecture 7. Proportional Hazards Model - Handling Ties and Survival Estimation Statistics Survival Analysis. Presented February 4, 2016

Introduction to repairable systems STK4400 Spring 2011

A multistate additive relative survival semi-markov model

Analysing geoadditive regression data: a mixed model approach

The Weibull Distribution

Model Based Clustering of Count Processes Data

Introduction to cthmm (Continuous-time hidden Markov models) package

Multistate models in survival and event history analysis

Part III. Hypothesis Testing. III.1. Log-rank Test for Right-censored Failure Time Data

Extensions of Cox Model for Non-Proportional Hazards Purpose

A Recursive Formula for the Kaplan-Meier Estimator with Mean Constraints

Parameters Estimation for a Linear Exponential Distribution Based on Grouped Data

MAS3301 / MAS8311 Biostatistics Part II: Survival

University of California, Berkeley

Transcription:

Multistate Modeling and Applications Yang Yang Department of Statistics University of Michigan, Ann Arbor IBM Research Graduate Student Workshop: Statistics for a Smarter Planet Yang Yang (UM, Ann Arbor) September 7, 00 / 4

Use of Multistate Models Focus on multiple events with inherent dependence stages of degradation, rehabilitation,... Offer flexible and general structure i.e. Disabled Twin Dead Alive Fail Healthy Both Alive Both Dead Dead Twin Dead Usual Failure Model Illness-death Model Twin Survival Model Markov models have been used extensively. especially in epidemiology, reliability and risk analysis. Semi-Markov models are becoming popular. Yang Yang (UM, Ann Arbor) September 7, 00 / 4

Time-to-Failure: From A Process Point of View Traditional time-to-failure analysis: focus merely on failure event Parametric models: Exponential, Weibull, Log-normal, Inverse-gaussian Nonparametric inference: Kaplan-Meier, Nelson-Aalen Regression modeling: Cox proportional hazards, Accelerated failure time From a process point of view: failure as the end point of some underlying process time to reach the absorbing state after moving among other states in the system Failure Figure: Phase-type distribution (Markov) more informative to model the entire process Yang Yang (UM, Ann Arbor) September 7, 00 3 / 4

An Illustrative Application Product: Home equity line of credit Natural states: current, -month-delinquent, -month-delinquent, 3-month-delinquent, default default time Y = 7 censoring time C = 8 MOB 3 4 5 6 7 8 9 0 3 4 5 6 7 8 A 0 0 0 0 6 0 0 0 0 0 0 0 0 0 0 0 8 56 A 0 0 0 0 4 3 0 0 0 0 5 0 7 A3 0 7 0 9 59 89 0 9 0 9 58 88 0 9 59 89 9 A4 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 Of interest: ) predicting time-to-default for various vintages (FICO, balance, HPI,...), or starting from x-month delinquent to default; ) do people who move from delinquency to current behave the same as people who are always current, etc. Yang Yang (UM, Ann Arbor) September 7, 00 4 / 4

Estimating the Default Distribution: Two approaches Consider a general semi-markov process J(t) for the multistate data. Let T d be the time-to-failure (absorption or default). Consider the hazard function λ d (t) = P(T d = t T d t) Two approaches to estimate λ d (t): i. Traditional estimator: i.e. Kaplan-Meier or Nelson-Aalen ignore the underlying multistate process ii. Exploit the underlying structure: estimate the initial distribution π, the form of the transition probability P(t) = {P ij (t)}, with P ij (t) P(J(t) = j J(0) = i) nonparametrically and use these to estimate λ d Yang Yang (UM, Ann Arbor) September 7, 00 5 / 4

Comparison of the Two Estimators Simulated data from a 5-state (Markov) model state 5 is absorbing Cumulative Hazard 0 3 4 5 6 Hazard 0.0 0. 0. 0.3 0.4 0.5 0 40 60 80 Time 0 0 40 60 80 Time Figure: ˆΛ P and ˆλ P (blue), ˆΛ T and ˆλ T with 95% pointwise confidence interval (black), and MLE (red) Traditional estimate ˆλ T is close initially but becomes highly variable in the tail (as the number at risk becomes small). Estimate ˆλ P that exploits the underlying process is more stable, closer to the MLE, and can be shown to be much more efficient. Yang Yang (UM, Ann Arbor) September 7, 00 6 / 4

Inference for Multistate Models with Panel Data In practice, multistate data are only observed at discrete times. Instead of exact transition times τ s and visited states i s, only the state information x k are observed at some discrete time t k. x 0 x t0 i x Z {( s, is); s,..., S} Y {( tk, xk); k,..., K} x i x x i x 3 3 3 t t t 3 Transition occurs Figure: Z latent complete history, Y observed censored history To make inference of parameters θ that characterizes J(t), likelihood function is: L(θ Y) = P(J(t 0) = x 0, J(t ) = x, J(t ) = x, J(t 3) = x 3) When J(t) is a Markov model, the likelihood can be evaluated (numerically). However, this is not true in general, even for semi-markov models. Yang Yang (UM, Ann Arbor) September 7, 00 7 / 4

Challenges with Panel Data 7 Underlying J(t) Hard to derive information about Z = {(τ s, i s); s =,..., S}: if x 0 =, x =, x = 3, x 3 = 4, the number of transition S is possibly! assumed that J(t) is acyclic/progressive (by removing the red transitions). The likelihood function L(θ Y) involves high dimensional integral. if x 0 =, x =, x = 3, x 3 = 4, and S = K, i k = x k, L(θ Y) = g(y)q (τ )q 3(τ )q 34(τ 3)dτ dτ dτ 3, where g(y) = {τ (t 0, t ], τ + τ (t, t ], τ + τ + τ 3 (t, t 3]}. Yang Yang (UM, Ann Arbor) September 7, 00 8 / 4

Likelihood-Based Inference via Stochastic Approximation and MCMC sampling For general semi-markov models, Observed (Incomplete) Y: L(θ Y) and h(θ, Y) θ log L(θ Y) are hard to evaluate. Unknown (Complete) Z: L(θ Z) and H(θ, Z) θ log L(θ Z) have explicit forms. Note h(θ, Y) = E Z θ,y H(θ, Z). If Z i p( θ, Y), h(θ, Y) can be empirically estimated by H(θ, Z) = m m i= H(θ, Z i ). To obtain MLE, we use stochastic approximation: Pick a positive sequence γ n s.t. n γn = and n γ n <. Update the estimate by θ n = θ n + γ n H(θ n, Z (n) ), () where Z (n) = {Z (n), Z(n),..., Z(n) m } and Z (n) i is sampled from p( θ n, Y) by MCMC sampling. We actually use an improved version of () proposed in Gu and Kong (PNAS, 998). Yang Yang (UM, Ann Arbor) September 7, 00 9 / 4

Sampling Z from p( θ, Y ) 3 x 0 x i 4 x δ denotes the sample path of the process. Note: The dimension of the unknown Z = (δ, τ ) can vary. Two sampling schemes: x ' ' i i x x i i 3 t t x x ' ' ' 0 t 3 t3 i Calculate the conditional probabilities of each path δ, given θ n and Y, then use Gibbs sampling to sample the multiple transition times τ ; ii Use reversible jump MCMC to handle the varying dimension directly, and sample δ and τ together. In our experience, reversible jump MCMC is computationally faster. x 3 Yang Yang (UM, Ann Arbor) September 7, 00 0 / 4

Illustration for a Markov model 3 3 34 3 4 4 4 λ.098.00.0.04.06.08 λ 3 0.85 0.830 0.835 0.840 0.845 0.850 0.855 0.860 λ 4 0.405 0.40 0.45 0.40 0 500 000 500 000 500 3000 0 500 000 500 000 500 3000 0 500 000 500 000 500 3000 Model: a 4-state progressive Markov model Data: 00 interval-censored observations λ 3 0.560 0.565 0.570 0.575 0.580 0.585 0.590 Iteration λ 4 0.605 0.60 0.65 0.60 0.65 Iteration λ 34.39.40.4.4 Iteration 0 500 000 500 000 500 3000 0 500 000 500 000 500 3000 0 500 000 500 000 500 3000 Iteration Iteration Iteration Figure: black dash MLE, red Conditional, green RJMCMC Yang Yang (UM, Ann Arbor) September 7, 00 / 4

Simulation Study F t F () t 3 () F () t 3 3 Figure: 3-state Semi-Markov Model n = 00 and t k t k = 0.3, with () F = LN(, ), F 3 = LN(, ), F 3 = LN(.5, ), p = 0.6 () F = W (0.8,.), F 3 = W (0.8,.), F 3 = W (, ), p = 0.4 (3) F = W (0.8,.), F 3 = LN(, ), F 3 = Exp(), p = 0.6 Conditional RJMCMC Model () 9099 6 Model () 84 687 Model (3) 57 675 Table: Computation time (second) of 3000 iterations in SA Yang Yang (UM, Ann Arbor) September 7, 00 / 4

Improvement in Inference and Prediction Model (3) with F = W (0.9,.), F 3 = LN(, ), F 3 = Exp(), p = 0.6 θ MLE = c(0.76,.7, 0.84, 0.99,.7, 0.5) θ Heurstic = c(.03,.6, 0., 0.85, 0.53, 0.) Hazard: >.0.5.0.5 Hazard: > 3 0.0 0.5.0.5.0 0 3 4 5 Time 0 3 4 5 Time Hazard: > 3 0.6 0.8.0..4.6 MRF: till absorbing.5.0.5 3.0 0 3 4 5 Time 0 3 4 5 Time Figure: red from MLE; black from heurstic estimates Yang Yang (UM, Ann Arbor) September 7, 00 3 / 4

Work in Progress Time-to-failure: from a process point of view Derived the nonparametric estimator ˆλ P that exploits the underlying process Comparison with the traditional estimator ˆλ T Asymptotic properties of ˆλ P Inference for semi-markov multistate model with panel data: Derived a computationally intense but general algorithm to estimate MLE Can be readily extended to include covariates (time non-varying), and non-progressive models are possible in principle Study of other parametric estimators based on pseduo-likelihood based on method of moments Semi-parametric and nonparametric estimator Non-semi-Markov Models Yang Yang (UM, Ann Arbor) September 7, 00 4 / 4