Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Bayesian statistics made simple by Allen Downey

Bayesian statistics made simple by Allen Downey

An introduction to Bayesian statistics using Python. Bayesian statistics are usually presented mathematically, but many of the ideas are easier to understand computationally. People who know some Python have a head start.
We will use material from Think Stats: Probability and Statistics for Programmers (O’Reilly Media), and Think Bayes, a work in progress at http://thinkbayes.com.

PyCon 2013

March 31, 2013
Tweet

More Decks by PyCon 2013

Other Decks in Programming

Transcript

  1. Goals At 4:20, you should be ready to: • Work

    on similar problems. • Learn more on your own.
  2. Think Stats This tutorial is based on my book, Think

    Stats: Probability and Statistics for Programmers. Published by O'Reilly Media and available under a Creative Commons license from thinkstats.com
  3. Think Bayes Also based on my new project, Think Bayes:

    Bayesian Statistics Made Simple Current draft available from thinkbayes.com
  4. Prerequisites Conditional probability p(A): the probability that A occurs. p(A|B):

    the probability that A occurs, given that B has occurred. Conjoint probability p(A and B) = p(A) p(B|A)
  5. Bayes's Theorem By definition of conjoint probability: p(A and B)

    = p(A) p(B|A) = (1) p(B and A) = p(B) p(A|B) Equate the right hand sides p(B) p(A|B) = p(A) p(B|A) (2) Divide by p(B) and ...
  6. Bayes's Theorem One way to think about BT: Bayes's Theorem

    is an algorithm to get from p (B|A) to p(A|B). Useful if p(B|A), p(A) and p(B) are easier than p (A|B). OR ...
  7. Diachronic interpretation H: Hypothesis D: Data Given p(H), the probability

    of the hypothesis before you saw the data. Find p(H|D), the probability of the hypothesis after you saw the data.
  8. A cookie problem Suppose there are two bowls of cookies.

    Bowl #1 has 10 chocolate and 30 vanilla. Bowl #2 has 20 of each. Fred picks a bowl at random, and then picks a cookie at random. The cookie turns out to be vanilla. What is the probability that Fred picked from Bowl #1? from Wikipedia
  9. Cookie problem H: Hypothesis that cookie came from Bowl 1.

    D: Cookie is vanilla. Given p(H), the probability of the hypothesis before you saw the data. Find p(H|D), the probability of the hypothesis after you saw the data.
  10. Diachronic interpretation p(H|D) = p(H) p(D|H) / p(D) p(H): prior

    p(D|H): conditional likelihood of the data p(D): total likelihood of the data
  11. Diachronic interpretation p(H|D) = p(H) p(D|H) / p(D) p(H): prior

    = 1/2 p(D|H): conditional likelihood of the data = 3/4 p(D): total likelihood of the data = 5/8
  12. Diachronic interpretation p(H|D) = (1/2)(3/4) / (5/8) = 3/5 p(H):

    prior = 1/2 p(D|H): conditional likelihood of the data = 3/4 p(D): total likelihood of the data = 5/8
  13. Partner up • If you don't have a working environment,

    find a neighbor who does. • Even if you do, try pair programming! • Take a minute to introduce yourself. • Questions? Ask your partner first (please).
  14. Icebreaker What was your first computer? What was your first

    programming language? What is the longest time you have spent finding a stupid bug?
  15. Start your engines 1. You downloaded bayes_pycon13.zip, right? 2. cd

    into that directory (or set PYTHON_PATH) 3. Start the Python interpreter. >>> from thinkbayes import Pmf
  16. Pmf from thinkbayes import Pmf # make an empty Pmf

    pmf = Pmf() # outcomes of a six-sided die for x in [1,2,3,4,5,6]: pmf.Set(x, 1)
  17. The framework 1) Build a Pmf that maps from each

    hypothesis to a prior probability, p(H). 2) Multiply each prior probability by the likelihood of the data, p(D|H). 3) Normalize, which divides through by the total likelihood, p(D).
  18. Update p(Vanilla | Bowl 1) = 30/40 p(Vanilla | Bowl

    2) = 20/40 pmf.Mult('Bowl 1', 0.75) pmf.Mult('Bowl 2', 0.5)
  19. The dice problem I have a box of dice that

    contains a 4-sided die, a 6-sided die, an 8-sided die, a 12-sided die and a 20-sided die. Suppose I select a die from the box at random, roll it, and get a 6. What is the probability that I rolled each die?
  20. Hypothesis suites A suite is a mutually exclusive and collectively

    exhaustive set of hypotheses. Represented by a Suite that maps hypothesis → probability.
  21. Suite class Suite(Pmf): "Represents a suite of hypotheses and their

    probabilities." def __init__(self, hypos): "Initializes the distribution." for hypo in hypos: self.Set(hypo, 1) self.Normalize()
  22. Dice from thinkbayes import Suite # start with equal priors

    suite = Suite([4, 6, 8, 12, 20]) suite.Print()
  23. Suite def Update(self, data): "Updates the suite based on data."

    for hypo in self.Values(): like = self.Likelihood(hypo, data) self.Mult(hypo, like) self.Normalize() self.Likelihood?
  24. Likelihood Outcome: 6 • What is the likelihood of this

    outcome on a six-sided die? • On a ten-sided die? • On a four-sided die? What is the likelihood of getting n on an m- sided die?
  25. Likelihood # hypo is the number of sides on the

    die # data is the outcome class Dice(Suite): def Likelihood(self, hypo, data): # write this method! Write your solution in dice.py
  26. Likelihood # hypo is the number of sides on the

    die # data is the outcome class Dice(Suite): def Likelihood(self, hypo, data): if hypo < data: return 0 else: return 1.0/hypo
  27. Dice # start with equal priors suite = Dice([4, 6,

    8, 12, 20]) # update with the data suite.Update(6) suite.Print()
  28. Dice Posterior distribution: 4 0.0 6 0.39 8 0.30 12

    0.19 20 0.12 More data? No problem...
  29. Dice for roll in [6, 4, 8, 7, 7, 2]:

    suite.Update(roll) suite.Print()
  30. Trains The trainspotting problem: • You believe that a freight

    carrier operates between 100 and 1000 locomotives with consecutive serial numbers. • You spot locomotive #321. • How many trains does the carrier operate? Modify train.py to compute your answer.
  31. Trains • If there are m trains, what is the

    chance of spotting train #n? • What does the posterior distribution look like? • What if we spot more trains? • Why did we do this example?
  32. A Euro problem "When spun on edge 250 times, a

    Belgian one- euro coin came up heads 140 times and tails 110. 'It looks very suspicious to me,' said Barry Blight, a statistics lecturer at the London School of Economics. 'If the coin were unbiased, the chance of getting a result as extreme as that would be less than 7%.' " From "The Guardian" quoted by MacKay, Information Theory, Inference, and Learning Algorithms.
  33. A Euro problem MacKay asks, "But do these data give

    evidence that the coin is biased rather than fair?" Assume that the coin has probability x of landing heads. (Forget that x is a probability; just think of it as a physical characteristic.)
  34. A Euro problem Estimation: Based on the data (140 heads,

    110 tails), what is x? Hypothesis testing: What is the probability that the coin is fair?
  35. Euro We can use the Suite template again. We just

    have to figure out the likelihood function.
  36. Likelihood # hypo is the prob of heads (1-100) #

    data is a string, either 'H' or 'T' class Euro(Suite): def Likelihood(self, hypo, data): # one more, please! Modify euro.py to compute your answer.
  37. Likelihood # hypo is the prob of heads (1-100) #

    data is a string, either 'H' or 'T' class Euro(Suite): def Likelihood(self, hypo, data): x = hypo / 100.0 if data == 'H': return x else: return 1-x
  38. Prior Start with something simple; we'll come back and review.

    Uniform prior: any value of x between 0% and 100% is equally likely.
  39. Update Suppose we spin the coin once and get heads.

    suite.Update('H') What does the posterior distribution look like? Hint: what is p(x=0% | D)?
  40. Update Suppose we spin the coin again, and get heads

    again. suite.Update('H') What does the posterior distribution look like?
  41. Update Suppose we spin the coin again, and get tails.

    suite.Update('T') What does the posterior distribution look like? Hint: what's p(x=100% | D)?
  42. Update After 10 spins, 7 heads and 3 tails: for

    outcome in 'HHHHHHHTTT': suite.Update(outcome)
  43. Update And finally, after 140 heads and 110 tails: evidence

    = 'H' * 140 + 'T' * 110 for outcome in evidence: suite.Update(outcome)
  44. Posterior • Now what? • How do we summarize the

    information in the posterior PMF?
  45. Posterior Given the posterior distribution, what is the probability that

    x is 50%? print pmf.Prob(50) And the answer is... 0.021 Hmm. Maybe that's not the right question.
  46. Posterior How about the most likely value of x? def

    MLE(pmf): prob, val = max((prob, val) for val, prob in pmf.Items()) return val And the answer is 56%.
  47. Posterior Or the expected value? def Mean(pmf): total = 0

    for val, prob in pmf.Items(): total += val * prob return total And the answer is 55.95%.
  48. Posterior The 5th percentile is 51. def Percentile(pmf, p): total

    = 0 for val, prob in sorted(pmf.Items()): total += prob if total >= p: return val
  49. Posterior The 5th percentile is 51. The 95th percentile is

    61. These values form a 90% credible interval. So can we say: "There's a 90% chance that x is between 51 and 61?"
  50. Bayesian response Yes, x is a random variable, Yes, (51,

    61) is a 90% credible interval, Yes, x has a 90% chance of being in it. However...
  51. The prior is subjective Remember the prior? We chose it

    pretty arbitrarily, and reasonable people might disagree. Is x as likely to be 1% as 50%? Given what we know about coins, I doubt it.
  52. Posterior What do you think the posterior distribution looks like?

    I was going to put an image here, but then I Googled "posterior". Never mind.
  53. Swamp the prior With enough data, reasonable people converge. But

    if any p(H i ) = 0, no data will change that.
  54. Swamp the prior Priors can be arbitrarily low, but avoid

    0. See wikipedia.org/wiki/ Cromwell's_rule "I beseech you, in the bowels of Christ, think it possible that you may be mistaken."
  55. Summary of estimation 1. Form a suite of hypotheses, H

    i . 2. Choose prior distribution, p(H i ). 3. Compute likelihoods, p(D|H i ). 4. Turn off brain. 5. Compute posteriors, p(H i |D).
  56. Hypothesis testing Remember the original question: "But do these data

    give evidence that the coin is biased rather than fair?" What does it mean to say that data give evidence for (or against) a hypothesis?
  57. Hypothesis testing D is evidence in favor of H if

    p(H|D) > p(H) which is true if p(D|H) > p(D|~H) or equivalently if p(D|H) / p(D|~H) > 1
  58. Hypothesis testing This term p(D|H) / p(D|~H) is called the

    likelihood ratio, or Bayes factor. It measures the strength of the evidence.
  59. Hypothesis testing F: hypothesis that the coin is fair B:

    hypothesis that the coin is biased p(D|F) is easy. p(D|B) is hard because B is underspecified.
  60. Bogosity Tempting: we got 140 heads out of 250 spins,

    so B is the hypothesis that x = 140/250. But, 1. Doesn't seem right to use the data twice. 2. By this process, almost any data would be evidence in favor of B.
  61. We need some rules 1. You have to choose your

    hypothesis before you see the data. 2. You can choose a suite of hypotheses, but in that case we average over the suite.
  62. Likelihood def AverageLikelihood(suite, data): total = 0 for hypo, prob

    in suite.Items(): like = suite.Likelihood(hypo, data) total += prob * like return total
  63. Hypothesis testing F: hypothesis that x = 50%. B: hypothesis

    that x is not 50%, but might be any other value with equal probability.
  64. Prior bias = Euro() for x in range(0, 101): if

    x != 50: bias.Set(x, 1) bias.Normalize()
  65. Bayes factor data = 140, 110 like_fair = AverageLikelihood(fair, data)

    like_bias = AverageLikelihood(bias, data) ratio = like_bias / like_fair
  66. Hypothesis testing And the answer is: p(D|B) = 2.6 ·

    10-76 p(D|F) = 5.5 · 10-76 Likelihood ratio is about 0.47. So this dataset is evidence against B.
  67. Fair comparison? • Modify the code that builds bias; try

    out a different definition of B and run again.
  68. Word problem for geeks ALICE: What did you get on

    the math SAT? BOB: 760 ALICE: Oh, well I got a 780. I guess that means I'm smarter than you. NARRATOR: Really? What is the probability that Alice is smarter than Bob?
  69. Assume, define, quantify Assume: each person has some probability, x,

    of answering a random SAT question correctly. Define: "Alice is smarter than Bob" means x a > x b . How can we quantify Prob { x a > x b } ?
  70. Be Bayesian Treat x as a random quantity. Start with

    a prior distribution. Update it. Compare posterior distributions.
  71. Likelihood def Likelihood(data, hypo): x = hypo score = data

    raw = self.exam.Reverse(score) yes, no = raw, self.exam.max_score - raw like = x**yes * (1-x)**no return like
  72. ProbBigger def ProbBigger(pmf1, pmf2): """Returns the prob that a value

    from pmf1 is greater than a value from pmf2."""
  73. ProbBigger def ProbBigger(pmf1, pmf2): """Returns the prob that a value

    from pmf1 is greater than a value from pmf2.""" Iterate through all pairs of values. Check whether the value from pmf1 is greater. Add up total probability of successful pairs.
  74. ProbBigger def ProbBigger(pmf1, pmf2): total = 0.0 for x1, p1

    in pmf1.Items(): for x2, p2 in pmf2.Items(): if x1 > x2: total += p1 * p2 return total
  75. Two points Posterior distribution is often the input to the

    next step in an analysis. Real world problems start (and end!) with modeling.
  76. Modeling • This result is based on the simplification that

    all SAT questions are equally difficult. • An alternative (in the book) is based on item response theory.
  77. Modeling • For most real world problems, there are several

    reasonable models. • The best choice depends on your goals. • Modeling errors probably dominate.
  78. Modeling Therefore: • Don't mistake the map for the territory.

    • Don't sweat approximations smaller than modeling errors. • Iterate.
  79. Dirichlet distribution • If we knew the number of species,

    we could estimate the prevalence of each.
  80. Hierarchical Bayes • For a hypothetical value of n, we

    can compute the likelihood of the data. • Using Bayes's Theorem, we can get the posterior distribution of n.
  81. Lions and tigers and bears • Suppose we have seen

    3 lions and 2 tigers and 1 bear. • There must be at least 3 species. • And up to 30?
  82. B1242 92 Corynebacterium 53 Anaerococcus 47 Finegoldia 38 Staphylococcus 15

    Peptoniphilus 14 Anaerococcus 12 unidentified 10 Clostridiales 8 Lactobacillales 7 Corynebacterium and 1 Partridgeinnapeartrium.
  83. What is this good for? These distributions know everything. •

    If I take more samples, how many more species will I see? • How many samples do I need to reach a given coverage?
  84. Case studies Think Bayes has 12 chapters. • Euro problem

    • SAT problem • Unseen species problem • Hockey problem • Paintball problem • Variability hypothesis • Kidney problem
  85. Case studies My students are working on case studies: •

    Fire (Bayesian regression) • Collecting whale snot • Poker • Bridge • Rating the raters • Predicting train arrival times
  86. Case studies Looking for more case studies. • Motivated by

    a real problem. • Uses the framework. • Extends the framework? • 5-10 page report. • Best reports included in the published version (pending contract).
  87. Remember • The Bayesian approach is a divide and conquer

    strategy. • You write Likelihood(). • Bayes does the rest.
  88. More reading MacKay, Information Theory, Inference, and Learning Algorithms Sivia,

    Data Analysis: A Bayesian Tutorial Gelman et al, Bayesian Data Analysis
  89. Where does this fit? Usual approach: • Analytic distributions. •

    Math. • Multidimensional integrals. • Numerical methods (MCMC).
  90. Where does this fit? Problem: • Hard to get started.

    • Hard to develop solutions incrementally. • Hard to develop understanding.
  91. My theory • Start with non-analytic distributions. • Use background

    information to choose meaningful priors. • Start with brute-force solutions. • If the results are good enough and fast enough, stop. • Otherwise, optimize (where analysis is one kind of optimization). • Use your reference implementation for regression testing.