BAYESIAN CURVE FITTING USING PIECEWISE POLYNOMIALS. Dariusz Biskup

Similar documents
XII.3 The EM (Expectation-Maximization) Algorithm

Excess Error, Approximation Error, and Estimation Error

Polynomial Regression Models

ITERATIVE ESTIMATION PROCEDURE FOR GEOSTATISTICAL REGRESSION AND GEOSTATISTICAL KRIGING

LECTURE :FACTOR ANALYSIS

Several generation methods of multinomial distributed random number Tian Lei 1, a,linxihe 1,b,Zhigang Zhang 1,c

Applied Mathematics Letters

Least Squares Fitting of Data

Outline. Prior Information and Subjective Probability. Subjective Probability. The Histogram Approach. Subjective Determination of the Prior Density

Denote the function derivatives f(x) in given points. x a b. Using relationships (1.2), polynomials (1.1) are written in the form

Least Squares Fitting of Data

System in Weibull Distribution

PROBABILITY AND STATISTICS Vol. III - Analysis of Variance and Analysis of Covariance - V. Nollau ANALYSIS OF VARIANCE AND ANALYSIS OF COVARIANCE

Markov Chain Monte Carlo (MCMC), Gibbs Sampling, Metropolis Algorithms, and Simulated Annealing Bioinformatics Course Supplement

8/25/17. Data Modeling. Data Modeling. Data Modeling. Patrice Koehl Department of Biological Sciences National University of Singapore

Markov Chain Monte-Carlo (MCMC)

1.3 Hence, calculate a formula for the force required to break the bond (i.e. the maximum value of F)

ANSWERS. Problem 1. and the moment generating function (mgf) by. defined for any real t. Use this to show that E( U) var( U)

Lecture Notes on Linear Regression

Durban Watson for Testing the Lack-of-Fit of Polynomial Regression Models without Replications

Nice plotting of proteins II

EFFECTS OF MAGNITUDE UNCERTAINTIES ON SEISMIC HAZARD ESTIMATES

Xiangwen Li. March 8th and March 13th, 2001

Computational and Statistical Learning theory Assignment 4

1 Definition of Rademacher Complexity

COS 511: Theoretical Machine Learning

What is LP? LP is an optimization technique that allocates limited resources among competing activities in the best possible manner.

Global Sensitivity. Tuesday 20 th February, 2018

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

arxiv: v2 [math.co] 3 Sep 2017

BAYESIAN AND NON BAYESIAN ESTIMATION OF ERLANG DISTRIBUTION UNDER PROGRESSIVE CENSORING

ASYMMETRIC TRAFFIC ASSIGNMENT WITH FLOW RESPONSIVE SIGNAL CONTROL IN AN URBAN NETWORK

Reliability estimation in Pareto-I distribution based on progressively type II censored sample with binomial removals

Bayesian estimation using MCMC approach based on progressive first-failure censoring from generalized Pareto distribution

The Multiple Classical Linear Regression Model (CLRM): Specification and Assumptions. 1. Introduction

Limited Dependent Variables

Motion Perception Under Uncertainty. Hongjing Lu Department of Psychology University of Hong Kong

A Robust Method for Calculating the Correlation Coefficient

1 Review From Last Time

Engineering Risk Benefit Analysis

Maximum Likelihood Estimation of Binary Dependent Variables Models: Probit and Logit. 1. General Formulation of Binary Dependent Variables Models

A Differential Evaluation Markov Chain Monte Carlo algorithm for Bayesian Model Updating M. Sherri a, I. Boulkaibet b, T. Marwala b, M. I.

4 Analysis of Variance (ANOVA) 5 ANOVA. 5.1 Introduction. 5.2 Fixed Effects ANOVA

Least squares cubic splines without B-splines S.K. Lucas

Multipoint Analysis for Sibling Pairs. Biostatistics 666 Lecture 18

A New Method for Estimating Overdispersion. David Fletcher and Peter Green Department of Mathematics and Statistics

PGM Learning Tasks and Metrics

MLE and Bayesian Estimation. Jie Tang Department of Computer Science & Technology Tsinghua University 2012

Statistical analysis using matlab. HY 439 Presented by: George Fortetsanakis

halftoning Journal of Electronic Imaging, vol. 11, no. 4, Oct Je-Ho Lee and Jan P. Allebach

CHAPT II : Prob-stats, estimation

CS 2750 Machine Learning. Lecture 5. Density estimation. CS 2750 Machine Learning. Announcements

Centroid Uncertainty Bounds for Interval Type-2 Fuzzy Sets: Forward and Inverse Problems

Analysis of Discrete Time Queues (Section 4.6)

Introducing Entropy Distributions

3.1 Expectation of Functions of Several Random Variables. )' be a k-dimensional discrete or continuous random vector, with joint PMF p (, E X E X1 E X

CS-433: Simulation and Modeling Modeling and Probability Review

Inference for the Rayleigh Distribution Based on Progressive Type-II Fuzzy Censored Data

Designing Fuzzy Time Series Model Using Generalized Wang s Method and Its application to Forecasting Interest Rate of Bank Indonesia Certificate

ON WEIGHTED ESTIMATION IN LINEAR REGRESSION IN THE PRESENCE OF PARAMETER UNCERTAINTY

{ In general, we are presented with a quadratic function of a random vector X

Negative Binomial Regression

Introduction to Regression

Probability and Random Variable Primer

Exam. Econometrics - Exam 1

The Geometry of Logit and Probit

RELIABILITY ASSESSMENT

ANSWERS CHAPTER 9. TIO 9.2: If the values are the same, the difference is 0, therefore the null hypothesis cannot be rejected.

PARAMETER ESTIMATION IN WEIBULL DISTRIBUTION ON PROGRESSIVELY TYPE- II CENSORED SAMPLE WITH BETA-BINOMIAL REMOVALS

Statistics for Economics & Business

Marginal Effects in Probit Models: Interpretation and Testing. 1. Interpreting Probit Coefficients

Properties of Least Squares

By M. O'Neill,* I. G. Sinclairf and Francis J. Smith

6. Stochastic processes (2)

Convexity preserving interpolation by splines of arbitrary degree

Hidden Markov Models

6. Stochastic processes (2)

/ n ) are compared. The logic is: if the two

Maximum Likelihood Estimation of Binary Dependent Variables Models: Probit and Logit. 1. General Formulation of Binary Dependent Variables Models

First Year Examination Department of Statistics, University of Florida

NUMERICAL DIFFERENTIATION

Supplementary material: Margin based PU Learning. Matrix Concentration Inequalities

Computation of Higher Order Moments from Two Multinomial Overdispersion Likelihood Models

Chapter 13: Multiple Regression

Chapter 1. Probability

Predictive Analytics : QM901.1x Prof U Dinesh Kumar, IIMB. All Rights Reserved, Indian Institute of Management Bangalore

ANALYSIS OF SIMULATION EXPERIMENTS BY BOOTSTRAP RESAMPLING. Russell C.H. Cheng

APPENDIX 2 FITTING A STRAIGHT LINE TO OBSERVATIONS

Using T.O.M to Estimate Parameter of distributions that have not Single Exponential Family

1. Inference on Regression Parameters a. Finding Mean, s.d and covariance amongst estimates. 2. Confidence Intervals and Working Hotelling Bands

Parametric fractional imputation for missing data analysis. Jae Kwang Kim Survey Working Group Seminar March 29, 2010

Comparison of Regression Lines

On the Construction of Polar Codes

SEMI-EMPIRICAL LIKELIHOOD RATIO CONFIDENCE INTERVALS FOR THE DIFFERENCE OF TWO SAMPLE MEANS

OPTIMIZATION BY SIMULATION METAMODELLING METHODS. Russell C.H. Cheng Christine S.M. Currie

Conjugacy and the Exponential Family

Lecture 13 APPROXIMATION OF SECOMD ORDER DERIVATIVES

4.3 Poisson Regression

An Experiment/Some Intuition (Fall 2006): Lecture 18 The EM Algorithm heads coin 1 tails coin 2 Overview Maximum Likelihood Estimation

Fermi-Dirac statistics

Transcription:

BAYESIAN CURVE FITTING USING PIECEWISE POLYNOMIALS Darusz Bskup 1. Introducton The paper presents a nonparaetrc procedure for estaton of an unknown functon f n the regresson odel y = f x + ε = N. (1) ( ), 1,..., It s assued that we have N observatons x, y of soe ndependent varable X and dependent varable Y. The observatons of Y are nfluenced by ε whch s a norally dstrbuted rando coponent wth ean zero and varance σ. The unknown regresson functon f can be approxated by a coposton of a certan nuber of low order polynoals defned on ntervals separated by the so called knot ponts. The Bayesan estaton of the regresson functon allows the nuber and the locaton of the knot ponts to be rando varables whch are estated usng the data. The paper presents several applcatons of ths ethod whch use Markov Chan Monte Carlo technques.. Pecewse polynoals The pecewse polynoal forula approxatng the functon f fro forula (1) s gven by (see [1]): where: k nuber of knots, ( a ) a = + ax,, l k l n n, + n,, () + n= = 1n= l n ( ) = β ( ) + β ( r ) f x x r x r knot ponts ndexed n the ascendng order wth the boundary knot r x, = k+ 1 n l order of the pecewse polynoal, r = x 1 l defnes the order of contnuty; f l = then the functon ay be dscontnuous, otherwse t s contnuous wth l 1 contnuous dervatves. and 1

The values for the l and l paraeters has to be fxed pror to estaton. If we set for exaple l = l = 3, then the functon () s called cubc polynoal splne. The Bayesan soluton to the proble of paraeter estaton does not requre pror specfcaton of the nuber and poston of the knots. These are both treated as rando varables whch are estated usng the avalable data. What s requred however s the specfcaton of the pror dstrbuton on the nuber of knots. The nuber of knots controls the tradeoff between the soothness of the functon and ts ft to the data ponts. The pror dstrbuton used has soe effect on the obtaned result. It sees however that t s not possble to autoatcally ft a curve wthout soe pror knowledge about how sooth t should be. There are two alternatve approaches to the poston of the knots. One of the requres that the knots are postoned on the data ponts x. The other whch s used n the paper has no such restrcton. 3. Pror and posteror dstrbuton In order to estate the odel () paraeters, t s necessary to specfy ther pror dstrbutons. The followng prors wll be assued: β ( n,..., l;,..., ) expectaton and standard devaton equal to 1 N ( ) noral ndependent prors for the n,, k 1 unfor prors on the nterval ( ) = = k paraeters wth ( n, ~, 1 ) β, ( ) r r + for the knot postons r = 1,.., k, gaa pror for the rando coponent varance: ~ G (,1;,1) σ, Posson pror for the nuber of knots k (the expectaton of ths pror wll vary between the exaples). Except for the pror on the nuber of knots, the assued prors are non-nforatve. The only paraeter whch requres soe subjectve judgent s the expectaton of the dstrbuton for k. It s also possble to assue a herarchcal pror for k (see [1]): k ~ Posson( λ ), ~ G( a, b) λ, where a and b are soe specfed constants controllng the uncertanty about the expected nuber of knots. Let us assue that the purpose s to approxate the value of the functon f at a certan pont x. We wll denote by Y the rando varable correspondng to the varable Y at the pont = { }, θ = ( σ, β, : n = n,..., l ; =,.., k ) x. Let also D ( x1, y1),...,( xn, yn ) ( kr, : 1,..., k) ϑ = = and M the dscrete set of possble odels. Then (see [4]): ( ) = (, ϑ ) = (, ϑ) ( ) p y D p y D p y D p ϑ D (3) ϑ M ϑ M The ter p( ϑ D) called odel probablty can be expressed n the followng way: ( ϑ ) ( ) p p( ϑ D) = p( Dθϑ, ) p( θϑ) dθ p D,,

where p( D ) s constant, p( ϑ ) s the pror probablty of a odel (n the paper t wll be a product of the Posson dstrbuton for k and unfor dstrbutons for r ), p( Dθ, ϑ ) s the lkelhood of the data D and p( θ ϑ ) s the pror dstrbuton of the paraeters θ n the current odel (here t wll be a product of norals and a gaa dstrbuton). The ter p( y D, ϑ ) n (3) s found usng the followng forula: p( y D, ϑ ) = p( y D, ϑθ, ) p( θd, ϑ) dθ, (4) where p( D, ) θ ϑ s the posteror dstrbuton of θ n odel ϑ and p( y,, ) noral wth expectaton f ( x ) and varance σ. D ϑ θ s It should be noted that the posteror dstrbuton of y s found usng not only the ost probable odel. It s expressed as an average of the predctons fro all the possble odels. Such procedure s called Bayesan odel averagng and s ore effcent than usng just one odel for predctons. The coputatons of the odel probablty are analytcally ntractable. Fndng the oents of the dstrbuton (3) s possble usng the reversble jup algorth (see [3] and [4] for detals). The algorth s pleented n the WINBUGS language (freely avalable fro http://www.rc-bsu.ca.ac.uk/bugs/wnbugs/contents.shtl). 4. Sulated exaples 4.1. Exaple 1 The frst functon to be approxated has the followng for (see []): A hundred pars of data ponts, ( ) ε ~ N ( ;,3) ( ) sn ( ) exp( 3 ), [,] f x = x + x x (5) ( x f x ε ) wth x equally spaced on the nterval + [-, ] and have been generated. The dataset s presented n Fg.. The data have been approxated wth a polynoal wth paraeters dstrbuton for the nuber of knots was chosen to be ( 5) l = l =. The pror Posson and Posson 1. 5 teratons of the reversble jup algorth have been perfored wth the frst 5 dscarded for the burn-n..6.4.. k saple: 1 4 6 k saple: 1 4 6 Fg. 1. Knot nuber dstrbuton for exaple 1 wth Posson(5) left and Posson(1) rght The dstrbuton for the nuber of knots s presented n Fg.. As we can see the ost probable value for the nuber of knots s 3 or dependng on the pror. It can also be.8.6.4.. () 3

seen that the axu nuber of knots does not exceed 5. Fg. 1 presents the ftted 5 Posson 1 curve for the Posson ( ) pror. The curve s practcally dentcal for the ( ) pror.,5 1,5 1,5 -,5 - -1,8-1,6-1,4-1, -1 -,8 -,6 -,4 -,,,4,6,8 1 1, 1,4 1,6 1,8-1 -1,5 - Fg.. Data for exaple 1 4.. Exaple The second functon s (see [1]): ( ) ( ) [ ] ( ) π ( ) ( ) f x = 4, 158sn 1+, 5 x+, 5 x 1 x, x, 1 (6) ( x f x ε) Fve hundreds pars of data ponts, ( ) ε ~ N ( ;1) wth x equally spaced on the nterval + [, 1] and have been generated. The dataset s presented n Fg. 4. The data have been approxated wth a polynoal wth paraeters dstrbuton for the nuber of knots was chosen to be ( 5) l = and l = 1. The pror Posson and Posson. 5 teratons of the reversble jup algorth have been perfored wth the frst 5 dscarded for the burn-n. The dstrbuton for the nuber of knots s presented n Fg. 3. As we can see the ode for the case ( s 47 and 8 for Posson 5. Fg. 4 presents the ftted curve for the ( ) Posson ) ( ) Posson () pror. The curve s practcally dentcal for the Posson ( 5) pror. As we can see the procedure s qute robust wth respect to the pror dstrbuton and for ths reason can be defned as alost autoatc. The ft for the exaple has also 4

been checked for the Posson() 1 pror. The result s that the curve s a lttle soother but also no serous dfference can be seen..1.75.5.5. k saple: 5 3 4 5 6.15.1.5. k saple: 5 19 3 4 Fg. 3. Knot nuber dstrbuton for exaple wth Posson() left and Posson(5) rght 15 1 5-5 -1-15 - Fg. 4. Data for exaple 4.3. Exaple 3,44,88,13,176,,64,38,35,396,44,484,58,57,616,66,74,748,79,836,88,94,968 The thrd exaple nvolves regresson of per capta GDP aganst lfe expectancy at brth. The dataset conssted of 174 observatons one for each country. As the curve s relatvely sooth wthout any dscontnutes the data have been approxated wth a polynoal wth paraeters was chosen to be l = l = Posson ( 5) and Posson ( 1). The pror dstrbuton for the nuber of knots. 5 teratons of the reversble jup algorth have been perfored wth the frst 5 dscarded for the burn-n. 5

The dstrbuton for the nuber of knots s presented n Fg. 5. As we can see the ost probable value for the nuber of knots s 1 or 4 dependng on the pror.. Fg. 6 presents the ftted curve for the Posson 1 pror. The curve s practcally dentcal for the Posson ( 5) pror..6.4.. k saple: 5 () 4 6 8.3..1. k saple: 95 5 1 15 Fg. 5. Knot nuber dstrbuton for exaple 3 wth Posson(1) left and Posson(5) rght 35 3 5 15 1 5 35 4 45 5 55 6 65 7 75 8 85 Fg. 6. Data for exaple 3 References: [1] Denson D.G.T., Mallck B.K., Sth A.F.M.: Autoatc Bayesan Curve Fttng, J.R. Statst. Soc. B, Vol. 6, 1998. [] DMatteo I., Genovese C.R., Kass R.E: Bayesan Curve Fttng wth Free-Knot Splnes, Boetrka 88, 1. [3] Green P.: Reversble Jup Markov Chan Monte Carlo Coputaton and Bayesan Model Deternaton. Boetrka 1995, 8, 711 73. [4] Lunn D. J., Best N., Whttaker J.: Generc reversble jup MCMC usng graphcal odels, Techncal Report EPH-5-1, Departent of Epdeology and Publc Health, Iperal College London, 5. Dr Darusz Bskup, Wroclaw Unversty of Econocs, Koandorska str. 118/1, 53-345 Wroclaw, Poland; darusz.bskup@ae.wroc.pl 6