Quiz 3. Model selection Overview Objectives determine the “choice” of model Modeling for forecasting Likelihood ratio test Akaike Information Criterion.

Slides:



Advertisements
Similar presentations
Regression Eric Feigelson Lecture and R tutorial Arcetri Observatory April 2014.
Advertisements

Econometric Modeling Through EViews and EXCEL
Brief introduction on Logistic Regression
Copyright © 2009 Pearson Education, Inc. Chapter 29 Multiple Regression.
Uncertainty and confidence intervals Statistical estimation methods, Finse Friday , 12.45–14.05 Andreas Lindén.
Hypothesis: It is an assumption of population parameter ( mean, proportion, variance) There are two types of hypothesis : 1) Simple hypothesis :A statistical.
Day 6 Model Selection and Multimodel Inference
Chap 9: Testing Hypotheses & Assessing Goodness of Fit Section 9.1: INTRODUCTION In section 8.2, we fitted a Poisson dist’n to counts. This chapter will.
Model Assessment and Selection
Model assessment and cross-validation - overview
Bayesian inference Gil McVean, Department of Statistics Monday 17 th November 2008.
G. Cowan Lectures on Statistical Data Analysis 1 Statistical Data Analysis: Lecture 10 1Probability, Bayes’ theorem, random variables, pdfs 2Functions.
458 Model Uncertainty and Model Selection Fish 458, Lecture 13.
458 Fitting models to data – II (The Basics of Maximum Likelihood Estimation) Fish 458, Lecture 9.
Course overview Tuesday lecture –Those not presenting turn in short review of a paper using the method being discussed Thursday computer lab –Turn in short.
Inference about a Mean Part II
Chapter 8: Inferences Based on a Single Sample: Tests of Hypotheses Statistics.
Aaker, Kumar, Day Seventh Edition Instructor’s Presentation Slides
BCOR 1020 Business Statistics
458 Fitting models to data – I (Sum of Squares) Fish 458, Lecture 7.
Copyright (c) 2004 Brooks/Cole, a division of Thomson Learning, Inc. Chapter 8 Tests of Hypotheses Based on a Single Sample.
INFERENTIAL STATISTICS – Samples are only estimates of the population – Sample statistics will be slightly off from the true values of its population’s.
Practical Model Selection and Multi-model Inference using R Presented by: Eric Stolen and Dan Hunt.
Aaker, Kumar, Day Ninth Edition Instructor’s Presentation Slides
Chapter 13: Inference in Regression
Overview Definition Hypothesis
Review of Statistical Inference Prepared by Vera Tabakova, East Carolina University.
Overview G. Jogesh Babu. Probability theory Probability is all about flip of a coin Conditional probability & Bayes theorem (Bayesian analysis) Expectation,
The Method of Likelihood Hal Whitehead BIOL4062/5062.
Single-Sample T-Test Quantitative Methods in HPELS 440:210.
Sullivan – Fundamentals of Statistics – 2 nd Edition – Chapter 11 Section 2 – Slide 1 of 25 Chapter 11 Section 2 Inference about Two Means: Independent.
BRIEF REVIEW OF STATISTICAL CONCEPTS AND METHODS.
Chapter 8 Introduction to Hypothesis Testing
01/20151 EPI 5344: Survival Analysis in Epidemiology Maximum Likelihood Estimation: An Introduction March 10, 2015 Dr. N. Birkett, School of Epidemiology,
Essential Statistics Chapter 131 Introduction to Inference.
Random Regressors and Moment Based Estimation Prepared by Vera Tabakova, East Carolina University.
VI. Evaluate Model Fit Basic questions that modelers must address are: How well does the model fit the data? Do changes to a model, such as reparameterization,
Lecture 4 Model Selection and Multimodel Inference.
Biostatistics Class 6 Hypothesis Testing: One-Sample Inference 2/29/2000.
Chapter 9 Fundamentals of Hypothesis Testing: One-Sample Tests.
1 Chapter 10: Introduction to Inference. 2 Inference Inference is the statistical process by which we use information collected from a sample to infer.
EMIS 7300 SYSTEMS ANALYSIS METHODS FALL 2005 Dr. John Lipp Copyright © Dr. John Lipp.
Introduction to the Practice of Statistics Fifth Edition Chapter 6: Introduction to Inference Copyright © 2005 by W. H. Freeman and Company David S. Moore.
Interval Estimation and Hypothesis Testing Prepared by Vera Tabakova, East Carolina University.
Maximum Likelihood Estimation Psych DeShon.
INTRODUCTION TO Machine Learning 3rd Edition
Practical Model Selection and Multi-model Inference using R Presented by: Eric Stolen and Dan Hunt.
Review of Statistical Inference Prepared by Vera Tabakova, East Carolina University ECON 4550 Econometrics Memorial University of Newfoundland.
Statistical Inference Statistical inference is concerned with the use of sample data to make inferences about unknown population parameters. For example,
Review of Statistical Inference Prepared by Vera Tabakova, East Carolina University.
Statistical Methods. 2 Concepts and Notations Sample unit – the basic landscape unit at which we wish to establish the presence/absence of the species.
Created by Erin Hodgess, Houston, Texas Section 7-1 & 7-2 Overview and Basics of Hypothesis Testing.
Hypothesis Testing. Statistical Inference – dealing with parameter and model uncertainty  Confidence Intervals (credible intervals)  Hypothesis Tests.
Model Selection Information-Theoretic Approach UF 2015 (25 minutes) Outline: Why use model selection Why use model selection AIC AIC AIC weights and model.
Review. Common probability distributions Discrete: binomial, Poisson, negative binomial, multinomial Continuous: normal, lognormal, beta, gamma, (negative.
© 2010 Pearson Prentice Hall. All rights reserved Chapter Hypothesis Tests Regarding a Parameter 10.
FIXETH LIKELIHOODS this is correct. Bayesian methods I: theory.
Delay-difference models. Readings Ecological Detective, p. 244–246 Hilborn and Walters Chapter 9.
Model Comparison. Assessing alternative models We don’t ask “Is the model right or wrong?” We ask “Do the data support a model more than a competing model?”
Model Comparison.
The 2 nd to last topic this year!!.  ANOVA Testing is similar to a “two sample t- test except” that it compares more than two samples to one another.
LOGISTIC REGRESSION. Purpose  Logistical regression is regularly used when there are only two categories of the dependent variable and there is a mixture.
Quiz 2.
Chapter 8: Inferences Based on a Single Sample: Tests of Hypotheses
Lecture 4 Model Selection and Multimodel Inference
Model Comparison.
Quantitative Methods in HPELS HPELS 6210
Essential Statistics Introduction to Inference
Lecture 4 Model Selection and Multimodel Inference
Lecture 4 Model Selection and Multimodel Inference
Presentation transcript:

Quiz 3

Model selection

Overview Objectives determine the “choice” of model Modeling for forecasting Likelihood ratio test Akaike Information Criterion (AIC) Bayes Information Criterion (BIC) Multi-model weighting

Readings Ecological Detective: pp Hobbs NT & Hilborn R (2006) Alternatives to statistical hypothesis testing in ecology: a guide to self teaching. Ecological Applications 16:5-19

Objectives of modelling and implications for model choice If we want to explore hypotheses: really complex models may be totally appropriate If we want to do forecasting: very simple models are usually best If we want to estimate uncertainty and do decision analysis: answer is less clear

Models for forecasting If we have too few parameters we can’t capture the underlying truth very well: error due to approximation If we have too many parameters we spend too much time fitting to minute vagaries in the data, and can’t estimate the parameters correctly: error due to estimation

Approximating the normal Sample a series of values from a normal distribution to create some simulated “data” Divide the data into equal intervals (“bins”) and use the mean within each bin to recreate the underlying distribution As we increase the number of bins (number of model parameters), there is a trade off between approximation and estimation error (This is a contrived example, the best model fit would be a normal distribution with parameters that are the mean and standard deviation of the data)

Approximation errorEstimation error Too few parameters to get a good fit to the truth. Too many parameters, model attempts to fit to every small deviation in the data. For more see Zucchini W (2000) An introduction to model selection. Journal of Mathematical Psychology 44:

Approximation error Estimation error Total error For more see Zucchini W (2000) An introduction to model selection. Journal of Mathematical Psychology 44:

Likelihood ratio test: nested models 4 parameters 3 parameters 2 parameters

Likelihood ratio test For nested models with more parameters, the log-likelihood (lnL) will usually be bigger (better fit). The likelihood ratio R (twice the difference between log-likelihood) follows a chi-square distribution with degrees of freedom equal to the difference in parameters between model A and model B.

Likelihood ratio test example B 1965 estimated (r, K, q, σ estimated) B 1965 = K (r, K, q, σ estimated) Modeln parslnL B 1965 = K B 1965 free Difference18.22 Likelihood ratio16.44 Chi-squared p-value In Excel, =CHISQ.DIST.RT(16.44,1) 2(lnL 1 -lnL 2 ) = 2( ) 9 Model selection.xlsx: sheet Likelihood ratio test

Problems with likelihood ratio Only valid for nested models so you cannot compare structurally different models Assumes the likelihoods are correct

Akaike Information Criterion (AIC) Origin of AIC: Akaike H (1973) Information theory and an extension of the maximum likelihood principle. In International Symposium on Information Theory. 2nd Edn. Edited by B.N. Petran and F. Csaàki. Akadèemiai Kiadi, Budapest, Hungary. pp

AIC corrected for small samples AIC c Since the correction term approaches 0 as n increases, I advise always using AIC c Number of observations (data points) Number of model parameters Negative log-likelihood Correction term Origin of AIC c : Hurvich CM & Tsai C (1989) Regression and time series model selection in small samples. Biometrika 76:

Why use AIC c ? Can be used to compare structurally different models (non-nested models) Used to find weights to put on different models (evidence for models)

AIC c weights How different the log likelihood of model i is from the best model The weight assigned each alternative model by AIC c

Comparing models using AIC The model with the lower AIC (or AIC c ) is the better model Rule of thumb from Burnham and Anderson (2002): – ΔAIC ≤ 2: substantial support (evidence) – 4 ≤ ΔAIC ≤ 7: considerably less support – ΔAIC ≥ 10: essentially no support Classic book: Burnham KP & Anderson DR (2002) Model selection and multi-model inference. Springer, Berlin, Heidelberg, New York.

AIC review Number of model parameters Model i Data (same data for all models) Number of data points (same for all models) The very best model The relative weight to give to competing models Can replace AIC with AIC c For model M i

Note about AIC For AIC and all related information criteria, -lnL is the minimum value of the log likelihood, as evaluated at the best choice of parameters for that particular model. (i.e. at the maximum likelihood, or MLE, estimate)

AIC example B 1965 estimated (r, K, q, σ estimated) B 1965 = K (r, K, q, σ estimated) 9 Model selection.xlsx Model M i Params p i NLLAICΔAICexp(-0.5Δ i )Weight w i B 1965 free B 1965 =K Model M i Params p i Data nNLLAIC c ΔAIC c exp(-0.5Δ i )Weight w i B 1965 free B 1965 =K Usually do not report these columns since only relative values are meaningful AIC c AIC

The LRSG model (Lagged recruitment, survival, and growth model) Lagged Beverton-Holt recruitment depends on biomass L years ago; L = years from egg deposition until available to fishing Starting conditions, unfished Catches in year t Recruitment assumed to be Beverton-Holt with steepness h Combined survival and somatic growth Hilborn & Mangel (2002) Ecological Detective, pp

Find the biomass where surplus production is maximized (B MSY ) Substitute that into the model to get MSY LRSG model

Model M i ParamsNLLAIC c ΔAIC c Weight LogisticB 1965 free LogisticB 1965 =K LSRGB 1965 free LSRGB 1965 =K Logistic LSRG Conclusion: Best model is logistic with initial stock size free parameter, it has the lowest AIC c 9 Model selection.xlsx

AIC and different likelihoods AIC can be used to compare models with different likelihood functions, if all the constants are included in the likelihoods Ken Burnham Always constant Constant if σ known not estimated Burnham KP & DR Anderson (2002) Model selection and multi-model inference. Springer, New York. Section 6.7, pp

Key points on model selection Appropriate model depends on use Methods – Likelihood ratio for nested models – AIC c for different models – AIC c for models with different likelihoods provided all the constants like sqrt(2π)are included Issues – Are likelihoods correct Likelihood ratio and AIC c are developed from theory about model fit, not uncertainty

Bayes Information Criterion, BIC Model M i ParamsNLLBICΔBIC LogisticB 1965 free LogisticB 1965 = K LSRGB 1965 free LSRGB 1965 = K Burnham KP & DR Anderson (2004) Multimodel inference: understanding AIC and BIC in model selection. Sociological Methods & Research 33: Number of model parameters Number of data points

Choosing a Criterion Options: R 2, LRT, other approaches Criteria: AIC, AICc, BIC, CAIC, CIC, DIC, EIC, FIC, GIC, HIC, ICOMP, JIC, KIC, NIC, OIC, PIC, QIC, QAIC, RIC, SIC, TIC, TAIC, WIC, YIC, ZIC, Bayes factor, Bayesian p-vals, Cross-validation Slide courtesy of Eric Ward, NOAA

Robustness and contradictory data

Robustness: Numerical Recipes: Chapter on “robust estimation” Contradictory data: Schnute JT & R Hilborn (1993) Analysis of contradictory data sources in fish stock assessment. CJFAS 50: Readings

Robustness In the real world, assumptions are not always met For instance, data may be mis-recorded, the wrong animal may be measured, the instrument may have failed, or some major assumption may have been wrong Outliers exist

Data 9 Robustness.xlsx

Data + contamination 9 Robustness.xlsx

Robust likelihoods

What is c?

Robust fit 9 Robustness.xlsx

Robust normal p = 1.00 p = 0.95p = 0.80