BOULDER WORKSHOP STATISTICS REVIEWED: LIKELIHOOD MODELS

Slides:



Advertisements
Similar presentations
Modeling of Data. Basic Bayes theorem Bayes theorem relates the conditional probabilities of two events A, and B: A might be a hypothesis and B might.
Advertisements

A. The Basic Principle We consider the multivariate extension of multiple linear regression – modeling the relationship between m responses Y 1,…,Y m and.
GENERAL LINEAR MODELS: Estimation algorithms
6-1 Introduction To Empirical Models 6-1 Introduction To Empirical Models.
Linear regression models
Ch11 Curve Fitting Dr. Deshi Ye
The General Linear Model. The Simple Linear Model Linear Regression.
1-1 Regression Models  Population Deterministic Regression Model Y i =  0 +  1 X i u Y i only depends on the value of X i and no other factor can affect.
Chapter 10 Simple Regression.
1 Chapter 3 Multiple Linear Regression Ray-Bing Chen Institute of Statistics National University of Kaohsiung.
Chapter 11 Multiple Regression.
Linear and generalised linear models
Basic Mathematics for Portfolio Management. Statistics Variables x, y, z Constants a, b Observations {x n, y n |n=1,…N} Mean.
Raw data analysis S. Purcell & M. C. Neale Twin Workshop, IBG Colorado, March 2002.
Linear regression models in matrix terms. The regression function in matrix terms.
Simple Linear Regression Analysis
Matrix Approach to Simple Linear Regression KNNL – Chapter 5.
Separate multivariate observations
Regression and Correlation Methods Judy Zhong Ph.D.
Lecture 5: Segregation Analysis I Date: 9/10/02  Counting number of genotypes, mating types  Segregation analysis: dominant, codominant, estimating segregation.
1 1 Slide © 2008 Thomson South-Western. All Rights Reserved Chapter 15 Multiple Regression n Multiple Regression Model n Least Squares Method n Multiple.
Univariate modeling Sarah Medland. Starting at the beginning… Data preparation – The algebra style used in Mx expects 1 line per case/family – (Almost)
Multiple Regression and Model Building Chapter 15 Copyright © 2014 by The McGraw-Hill Companies, Inc. All rights reserved.McGraw-Hill/Irwin.
Chapter 13 Multiple Regression
Lecture 3: Statistics Review I Date: 9/3/02  Distributions  Likelihood  Hypothesis tests.
Multiple Regression. Simple Regression in detail Y i = β o + β 1 x i + ε i Where Y => Dependent variable X => Independent variable β o => Model parameter.
Copyright © 2011 by The McGraw-Hill Companies, Inc. All rights reserved. McGraw-Hill/Irwin Simple Linear Regression Analysis Chapter 13.
Tutorial I: Missing Value Analysis
Mx modeling of methylation data: twin correlations [means, SD, correlation] ACE / ADE latent factor model regression [sex and age] genetic association.
Lecture 23: Quantitative Traits III Date: 11/12/02  Single locus backcross regression  Single locus backcross likelihood  F2 – regression, likelihood,
1 1 Slide The Simple Linear Regression Model n Simple Linear Regression Model y =  0 +  1 x +  n Simple Linear Regression Equation E( y ) =  0 + 
Nonparametric Statistics
MathematicalMarketing Slide 5.1 OLS Chapter 5: Ordinary Least Square Regression We will be discussing  The Linear Regression Model  Estimation of the.
The “Big Picture” (from Heath 1995). Simple Linear Regression.
LOGISTIC REGRESSION. Purpose  Logistical regression is regularly used when there are only two categories of the dependent variable and there is a mixture.
STA302/1001 week 11 Regression Models - Introduction In regression models, two types of variables that are studied:  A dependent variable, Y, also called.
I. Statistical Methods for Genome-Enabled Prediction of Complex Traits OUTLINE THE CHALLENGES OF PREDICTING COMPLEX TRAITS ORDINARY LEAST SQUARES (OLS)
Lecture 2 Survey Data Analysis Principal Component Analysis Factor Analysis Exemplified by SPSS Taylan Mavruk.
Applied statistics Usman Roshan.
Extended Pedigrees HGEN619 class 2007.
Nonparametric Statistics
The simple linear regression model and parameter estimation
Why Model? Make predictions or forecasts where we don’t have data.
Probability Theory and Parameter Estimation I
Regression 10/29.
Correlation, Regression & Nested Models
Re-introduction to openMx
CH 5: Multivariate Methods
Correlation – Regression
Regression.
MRC SGDP Centre, Institute of Psychiatry, Psychology & Neuroscience
Univariate modeling Sarah Medland.
Chapter 9 Hypothesis Testing.
Nonparametric Statistics
6-1 Introduction To Empirical Models
Modelling data and curve fitting
Regression Models - Introduction
Why general modeling framework?
Pak Sham & Shaun Purcell Twin Workshop, March 2002
OVERVIEW OF LINEAR MODELS
(Re)introduction to Mx Sarah Medland
Interval Estimation and Hypothesis Testing
Sarah Medland faculty/sarah/2018/Tuesday
Simple Linear Regression
OVERVIEW OF LINEAR MODELS
Multivariate Linear Regression
Power Calculation for QTL Association
3.2. SIMPLE LINEAR REGRESSION
Introduction to Regression
Presentation transcript:

BOULDER WORKSHOP STATISTICS REVIEWED: LIKELIHOOD MODELS Andrew C. Heath

PRE-HISTORY (STATISTICS 101) Binomial distribution – gives the probabilities of various possible numbers of ‘successful’ outcomes in a fixed number of discrete trials, where all trials have the same probability of success. Probability that X is equal to a particular value x (x= 0, 1, 2…n) is given by Useful for genetics (e.g. transmission versus non-transmission of an allele)!

LIKELIHOOD Focus on probability (“likelihood”) of the data, as a function of model parameters. E.g. Sham (1998) uses Neel & Schull (1954) data on opalescent dentine in random sample of 112 offspring of an affected parent, found 52 affected, 60 normal. Compatible with hypothesis of rare autosomal dominant? Does the observed population (.464) differ from expected proportion of 0.5? Likelihood function for segregation ratio p is

MAXIMUM LIKELIHOOD ESTIMATION = Find the maximum value of the likelihood function in the range 0  p 1. In more difficult problems, usual to maximize the log-likelihood, since computationally more convenient, & this also maximizes the likelihood. In this simple case, maximum likelihood estimate (MLE) of p, = r/n. For the opalescent dentine data, ignoring the constant term involving n, r which does not vary as a function of p, log-likelihood function is ln L(p) = 52 ln(p) + 60 ln(1–p)

Figure 2.1 Log-likelihood function of the segregation ratio for the opalescent dentine data (from Sham, 1998) p

LIKELIHOOD-RATIO STATISTIC Likelihood ratio statistic: twice the difference between the log likelihood of the data at the MLE (i.e. ), L1, and the log likelihood of the data at the hypothesized value of 0.5, L0: 2(ln L1 – ln L0). For the segregation ratio example, = 0.57 for the opalescent dentine example. Likelihood-ratio statistic in this case is distributed as chi-square on one degree of freedom, hence non-significant.

MATRIX ALGEBRA BASICS A–1 Inverse of A AT or A' Transpose of A |A| Determinant of A A  B A postmultiplied by B: r x c c x r conformable for multiplication since number of columns of A = number of rows of B. Resulting matrix is r x r. Tr (A) Trace of matrix A

HISTORY (MX INTRODUCTORY WORKSHOP) Maximum-likelihood estimation using linear covariance structure models, e.g. fitting models to twin data: Let p be the number of observed variables, the expected covariance matrix be E, and the expected vector of means be , where E and  are functions of q free parameters to be estimated from the data. Let x1, x2…xn denote to observed variables. Assuming that the observed variables follow a multivariate normal distribution, the log-likelihood of the observed data is given by This is the formula used for maximum likelihood model-fitting to raw continuous data, assuming a multivariate normal distribution. (Often, 2 ln L, is estimated). Requires that we provide: (a) model for expected covariance matrix – e.g. in terms of additive genetic, shared and non-shared environmental variance components – that will vary as a function of relationship; (b) model for expected means – in simplest applications we might estimate a separate mean that might differ by gender, or possibly by twin pair zygosity.

EXAMPLE: MZ Pairs: E =  = [ m , m] T DZ Pairs:  = [ m , m]T Where m is estimate of population mean, VA, VC, VE and are additive genetic, shared environmental and non-shared environmental variances, all estimated jointly from the data. Compare e.g. ln L1 for VA VC VE model ln L0 for VC VE model 2(ln L1 – ln L0) distributed as chi-square on one degree of freedom as before.

HISTORY (MX INTRODUCTORY WORKSHOP) - II In practice most applications in the MX Introductory Workshop fitted models to summary covariance matrices. We can simplify (e.g. Sham, p. 238): (1) (2) Where m is the vector of sample means of the p observed variables, S is the observed covariance matrix. For the simple applications in the Introductory Workshop, we had no predictions for the means structure, so we can saturate that component of the model (i.e. estimate a separate mean for every observed mean), equivalent to deleting the term (m)T E-1 (m) in (2). Thus the log-likelihood of the observed data becomes (3) Under a saturated model, which equates every element of E to the corresponding element of S (i.e. a perfect fit model) we have for the log-likelihood

HISTORY (MX INTRODUCTORY WORKSHOP) - II Thus the likelihood-radio test of the fitted model against the saturated model becomes (4) For multiple group problem, sum over groups.

Analysis of Australian BMI data-young female MZ twins pairs-MX DYI version Analysis of Australian BMI data-young female MZ twins pairs-MX DIY version ! DA NG=1 NI=2 NO=0 begin matrices; A LO 1 1 FR ! Additive genetic variance C LO 1 1 FR ! Shared environmental variance E LO 1 1 FR ! Non-shared environmental variance M FU 2 2 ! This will be observed MZ covariance matrix D FU 2 2 ! This will be observed DZ covariance matrix g fu 1 1 ! coefficient of 0.5 for DZ pairs n fu 1 1 ! sample size for MZ pairs (female in this illustration) k fu 1 1 ! sample size for DZ pairs (female in this illustration) p fu 1 1 ! order of matrices (i.e. number of variables =2 in this case) end matrices; mat g 0.5 mat p 2 mat m 0.7247 0.5891 0.5891 0.7915 mat n 532 mat d 0.7786 0.2461 0.2461 0.8365 mat k 326 mat a 0.25 mat c 0.25 mat e 0.25

Analysis of Australian BMI data-young female MZ twins pairs-MX DYI version (ctd) BEGIN ALGEBRA; t=n+k; ! total sample size U=(A+C+E | A+C _ A+C | A+C+E); ! Expected MZ covariance matrix V=(A+C+E | g*A+C _ g*A+C | A+C+E); ! Expected DZ covariance matrix H=n*(\ln (\det(U))-\ln (\det(M)) + \tr((U~ *M))-p); ! fit-function for MZ group J=k*(\ln (\det(V))-\ln (\det(D)) + \tr((V~ *D))-p); ! fit-function for DZ group F=h+j; END ALGEBRA; bo 0.01 1.0 e(1,1) bo 0.0 1.0 c(1,1) a(1,1) CO F ; option user df=6 end

PRE-HISTORY (STATISTICS 101) - II LINEAR REGRESSION: requires weaker assumptions than linear covariance structure models. Does not assume multivariate normal distribution, only homoscedastic residuals. Flexible for handling selective sampling schemes where we oversample extreme values of predictor variable(s). We can fit linear regression models by maximum likelihood e.g. using MX.

HISTORY! MX INTRODUCTORY WORKSHOP - III “Definition variables” – an option in MX when fitting to raw data, which allows us to model effects of some variables as fixed effects, modeling their contribution to expected means. Simple example: controlling for linear or polynomial regression of a quantitative measure on age. Don’t want to model covariance structure with age (which probably has rectangular distribution!) Definition variables = variables whose values may vary from individual to individual, that can be read into matrices Important example: genotypes at a given locus or set of loci.