Bayesian Statistics, Modeling & Reasoning What is this course about? P548: Intro Bayesian Stats with Psych Applications Instructor: John Miyamoto 01/04/2016: Lecture 01-1 Note: This Powerpoint presentation may contain macros that I wrote to help me create the slides. The macros aren’t needed to view the slides. You can disable or delete the macros without any change to the presentation. Outline • What is Bayesian inference? • Why are Bayesian statistics, modeling & reasoning relevant to psychology? • What is Psych 548 about? • Explain Psych 548 website • Intro to R • Intro to RStudio • Intro to the R to BUGS interface Psych 548, Miyamoto, Win '16 Lecture probably ends here 2 Bayes Rule – What Is It? • Reverend Thomas Bayes, 1702 – 1761 English Protestant minister & mathematician • Bayes Rule is fundamentally important to: ♦ Bayesian statistics ♦ Bayesian decision theory ♦ Bayesian models in psychology P Data|Hypothesis P(Hypothesis) P(Hypothesis|Data) = P(Data) P(Data) n P Data | Hypothesisi P Hypothesisi i 1 Psych 548, Miyamoto, Win '16 Bayes Rule – Why Is It Important? 3 Bayes Rule – Why Is It Important? • Bayes Rule is the optimal way to update the probability of hypotheses given data. • The concept of "Bayesian reasoning“: 3 related concepts ♦ Concept 1: Bayesian inference is a model of optimal learning from experience. ♦ Concept 2: Bayesian decision theory describes optimal strategies for taking actions in an uncertain environment. Optimal gambling strategies. ♦ Concept 3: Bayesian reasoning represents the uncertainty of events as probabilities in a mathematical calculus. • Concepts 1, 2 & 3 are all consistent with the use of the term, "Bayesian", in modern psychology. Psych 548, Miyamoto, Win '16 Bayesian Issues in Psychology 4 Bayesian Issues in Psychological Research • Does human reasoning about uncertainty conform to Bayes Rule? Do humans reason about uncertainty as if they are manipulating probabilities? ♦ These questions are posed with respect to infants & children, as well as adults. • Do neural information processing systems (NIPS) incorporate Bayes Rule? • Do NIPS model uncertainties as if they are probabilities. Psych 548, Miyamoto, Win '16 Four Roles for Bayesian Reasoning in Psychology Research 5 Four Roles for Bayesian Reasoning in Psychology 1. Bayesian statistics: Analyzing data ♦ E.g., is the slope of the regression of grades on IQ the same for boys as for girls? ♦ E.g., are there group differences in an analysis of variance? Psych 548, Miyamoto, Win '16 Four Roles …. (Continued) 6 Four Roles for Bayesian Reasoning in Psychology 1. Bayesian statistics: Analyzing data 2. Bayesian decision theory – a theory of strategic action. How to gamble if you must. 3. Bayesian modeling of psychological processes 4. Bayesian reasoning – Do people reason as if they are Bayesian probability analysts? (At macro & neural levels) ♦ Judgment and decision making – This is a major issue. ♦ Human causal reasoning – is it Bayesian or quasi-Bayesian? ♦ Modeling neural decision making – many proposed models have a strong Bayesian flavor. Psych 548, Miyamoto, Win '16 Four Roles …. (Continued) 7 Four Roles for Bayesian Reasoning in Psychology 1. Bayesian statistics: Analyzing data 2. Bayesian decision theory – a theory of strategic action. How to gamble if you must. 3. Bayesian modeling of psychological processes 4. Bayesian reasoning – Do people reason as if they are Bayesian probability analysts? (At macro & neural levels) Psych 548: Focus on Topics (1) and (3). Include a little bit of (4). Psych 548, Miyamoto, Win '16 Graphical Representation of Psych 548 Focus on Stats/Modeling 8 Graphical Representation of Psych 548 Psych 548 Bayesian Statistics & Modeling: R & JAGS Psych 548, Miyamoto, Win '16 Bayesian Models in Cognitive Psychology & Neuroscience Graph & Text Showing the History of S, S-Plus & R 9 Brief History of S, S-Plus, & R • S – open source statistics program created by Bell Labs (1976 – 1988 – 1999) Ancestry of R • S-Plus – commercial statistics program, S refinement of S (1988 – present) • R – free open source statistics program (1997 – present) ♦ currently the standard computing framework for statisticians worldwide Many contributors to its development ♦ Excellent general computation. Powerful & flexible. Great graphics. Multiplatform: Unix, Linux, Windows, Mac User must like programming ♦ ♦ ♦ Psych 548, Miyamoto, Win '16 S-Plus R BUGS, WinBUGS, OpenBUGS, JAGS 10 BUGS, WinBUGS, OpenBUGS & JAGS • Gibbs Sampling & Metropolis-Hastings Algorithm Two algorithms for sampling from a hard-to-evaluate probability distribution. “BUGS” includes all of these. • BUGS – Bayesian inference Under Gibbs Sampling (circa 1995) • WinBUGS - Open source (circa 1997) ♦ Windows only • OpenBUGS – Open source (circa 2006) ♦ Mainly Windows. Runs within a virtual Windows machine on a Mac. • JAGS – Open source (circa 2007) ♦ Multiplatform: Windows, Mac, Linux • STAN – Open source (circa 2012) ○ Multiplatform: Windows, Mac, Linux Psych 548, Miyamoto, Win '16 Basic Structure of Bayesian Computation with R & OpenBUGS 11 Basic Structure of Bayesian Computation rjags runjags R JAGS rjags functions data preparation rjags functions analysis of results BRugs R2WinBUGS rstan R Psych 548, Miyamoto, Win '16 BRugs functions Brugs functions Computes approximation to the posterior distribution. Includes diagnostics. OpenBUGS/ WinBUGS/ Stan Outline of Remainder of the Lecture: Course Outline & General Information 12 RStudio • Run RStudio • Run R from within RStudio Psych 548, Miyamoto, Win '16 13 Remainder of This Lecture • Take 5 minute break • Introduce selves • Psych 548: What will we study? • Briefly view the Psych 548 webpage. • Introduction to the computer facility in CSSCR. • Introduction to R, BUGS (OpenBUGS & JAGS), and RStudio Psych 548, Miyamoto, Win '16 5 Minute Break 14 5 Minute Break • Introduce selves upon return Psych 548, Miyamoto, Win '16 Course Goals 15 Course Goals • Learn the theoretical framework of Bayesian inference. • Achieve competence with R, OpenBUGS and JAGS. • Learn basic Bayesian statistics ♦ Learn how to think about statistical inference from a Bayesian standpoint. ♦ Learn how to interpret the results of a Bayesian analysis. ♦ Learn basic tools of Bayesian statistical inference - testing for convergence, making standard plots, examing samples from a posterior distribution. --------------------------------------------------------------Secondary Goals ♦ Bayesian modeling in psychology ♦ Understand arguments about Bayesian reasoning in the psychology of reasoning. The pros and cons of the heuristics & biases movement. Psych 548, Miyamoto, Win '16 Kruschke Textbook 16 Kruschke, Doing Bayesian Data Analysis Kruschke, J. K. (2014). Doing bayesian data analysis, second edition: A tutorial with R, JAGS, and Stan. Academic Press. • Excellent textbook – worth the price ($90 from Amazon) • Emphasis on classical statistical test problems from a Bayesian perspective. Not so much modeling per se. ♦ Binomial inference problems, anova problems, linear regression problems. Computational Requirements • R & JAGS (or OpenBUGS) • A programming editor like Rstudio is useful. Psych 548, Miyamoto, Win '16 Chapter Outline of Kruschke Textbook 17 Kruschke, Doing Bayesian Data Analysis • Ch 1 – 4: Basic probability background (pretty easy) • Ch 5 – 8: Bayesian inference with simple binomial models ♦ Conjugate priors, Gibbs sampling & Metropolis-Hastings algorithm ♦ OpenBUGS or JAGS • Ch 9 – 12: Bayesian approach to hierarchical modeling, model comparison, & hypothesis testing. • Ch 13: Power & sample size (omit ) • Ch 14: Intro generalized linear model • Ch 15 – 17: Intro linear regression • Ch 18 – 19: Oneway & multifactor anova • Ch 20 – 22: Categorical data analysis, logistic regression, probit regression, poisson regression Psych 548, Miyamoto, Win '16 Lee & Wagenmakers, Bayesian Graphical Modeling 18 Bayesian Cognitive Modeling Lee, M. D., & Wagenmakers, E. J. (2014). Bayesian cognitive modeling: A practical course. Cambridge University Press. ♦ Michael Lee: ♦ E. J. Wagenmaker: http://users.fmg.uva.nl/ewagenmakers/BayesCourse/BayesBook.html ♦ Equivalent Matlab & R code for book are available at the Psych 548 website and at Lee or Wagenmaker's website. http://www.socsci.uci.edu/~mdlee/bgm.html • Emphasis is on Bayesian models of psychological processes rather than on methods of data analysis. Lots of examples. Psych 548, Miyamoto, Win '16 Chapters in Lee & Wagenmakers 19 Table of Contents in Lee & Wagenmakers Psych 548:, Miyamoto, Win ‘16 Computer Setup in CSSCR 20 CSSCR Network & Psych 548 Webpage • Click on /Start /Computer. The path & folder name for your Desktop is: C:\users\NetID\Desktop (where "NetID" refers to your NetID) • Double click on MyUW on your Desktop. Find Psych 548 under your courses and double click on the Psych 548 website. Is this information obsolete? • Download files that are needed for today's class. Save these files to ♦ C:\users\NetID\Desktop Note that Ctrl-D takes you to your Desktop. • Run R or RStudio. Psych 548, Miyamoto, Win '16 Psych 548 Website - END 21 Psych 548 Website • Point out where to download the material for today’s class • Point out pdf’s for the textbooks. Psych 548, Miyamoto, Win '16 NEXT: Time Permitting ...... 22 General Characteristics of Bayesian Inference • The decision maker (DM) is willing to specify the prior probability of the hypotheses of interest. • DM can specify the likelihood of the data given each hypothesis. • Using Bayes Rule, we infer the probability of the hypotheses given the data Psych 548, Miyamoto, Win '16 Comparison Between Bayesian & Classical Stats - END 23 How Does Bayesian Stats Differ from Classical Stats? Bayesian: Common Aspects Classical: Common Aspects • Statistical Models • Statistical Models • Credible Intervals – sets of • Confidence Intervals – which parameter values are tenable after viewing the data. parameters that have high posterior probability Bayesian: Divergent Aspects Classical: Divergent Aspects • Given data, compute the full posterior probability distribution over all parameters • No prior distributions in general, so this idea is meaningless or selfdeluding. • Generally null hypothesis testing is nonsensical. • Null hypothesis te%sting • Posterior probabilities are meaningful; p-values are half-assed. • MCMC approximations are sometimes useful but not for computing posterior distributions. • MCMC approximations to posterior distributions. Psych 548, Miyamoto, Win '16 • P-values Sequential Presentation of the Common & Divergent Aspects 24 How Does Bayesian Stats Differ from Classical Stats? Bayesian: Common Aspects Classical: Common Aspects • Statistical Models • Statistical Models • Credible Intervals – sets of parameters that have high posterior probability • Confidence Intervals – which parameter values are tenable after viewing the data. Bayesian: Divergent Aspects Classical: Divergent Aspects • Given data, compute the full posterior probability distribution over all parameters • No prior distributions in general, so this idea is meaningless or selfdeluding. • Generally null hypothesis testing is nonsensical. • Null hypothesis te%sting • Posterior probabilities are meaningful; p-values are half-assed. • MCMC approximations are sometimes useful but not for computing posterior distributions. • MCMC approximations to posterior distributions. Psych 548, Miyamoto, Win '16 • P-values END 25 How Does Bayesian Stats Differ from Classical Stats? Bayesian: Common Aspects Classical: Common Aspects • Statistical Models • Statistical Models • Credible Intervals – sets of parameters that have high posterior probability • Confidence Intervals – which parameter values are tenable after viewing the data. Bayesian: Divergent Aspects Classical: Divergent Aspects • Given data, compute the full posterior probability distribution over all parameters • No prior distributions in general, so this idea is meaningless or self-deluding. • Generally null hypothesis testing is nonsensical. • Null hypothesis testing • Posterior probabilities are meaningful; p-values are half-assed. • MCMC approximations are sometimes useful but not for computing posterior distributions. • MCMC approximations to posterior distributions. Psych 548, Miyamoto, Win '16 • P-values END 26 How Does Bayesian Stats Differ from Classical Stats? Bayesian: Common Aspects Classical: Common Aspects • Statistical Models • Statistical Models • Credible Intervals – sets of parameters that have high posterior probability • Confidence Intervals – which parameter values are tenable after viewing the data. Bayesian: Divergent Aspects Classical: Divergent Aspects • Given data, compute the full posterior probability distribution over all parameters • No prior distributions in general, so this idea is meaningless or selfdeluding. • Generally null hypothesis testing is nonsensical. • Null hypothesis testing • Posterior probabilities are meaningful; p-values are half-assed. • MCMC approximations are sometimes useful but not for computing posterior distributions. • MCMC approximations to posterior distributions. Psych 548, Miyamoto, Win '16 • P-values END 27