Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Features
Speaker Deck
PRO
Sign in
Sign up for free
Search
Search
Statistical Thinking for Data Science
Search
Chris Fonnesbeck
February 08, 2015
Science
5
1.1k
Statistical Thinking for Data Science
PyTennessee 2015 Keynote Address
Chris Fonnesbeck
February 08, 2015
Tweet
Share
More Decks by Chris Fonnesbeck
See All by Chris Fonnesbeck
Structured Decision-making and Adaptive Management For The Control Of Infectious Disease
fonnesbeck
3
110
Estimating Microbial Diversity
fonnesbeck
0
110
Bayesian Statistical Analysis: A Gentle Introduction
fonnesbeck
4
620
Other Decks in Science
See All in Science
大規模言語モデルの開発
chokkan
PRO
85
43k
20分で分かる Human-in-the-Loop 機械学習におけるアノテーションとヒューマンコンピューターインタラクションの真髄
hurutoriya
5
2.8k
トラブルがあったコンペに学ぶデータ分析
tereka114
2
1.4k
インフラだけではない MLOps の話 @事例でわかるMLOps 機械学習の成果をスケールさせる処方箋 発売記念
icoxfog417
PRO
2
720
Healthcare Innovation through Business Entrepreneurship
clintwinters
0
200
Visual Analytics for R&D Intelligence @Funding the Commons & DeSci Tokyo 2024
hayataka88
0
140
Iniciativas independentes de divulgação científica: o caso do Movimento #CiteMulheresNegras
taisso
0
1k
非同期コミュニケーションの構造 -チャットツールを用いた組織における情報の流れの設計について-
koisono
0
210
創薬における機械学習技術について
kanojikajino
16
5k
ACL読み会2024@名大 REANO: Optimising Retrieval-Augmented Reader Models through Knowledge Graph Generation
takuma_matsubara
0
150
論文紹介: PEFA: Parameter-Free Adapters for Large-scale Embedding-based Retrieval Models (WSDM 2024)
ynakano
0
220
拡散モデルの原理紹介
brainpadpr
3
6k
Featured
See All Featured
Agile that works and the tools we love
rasmusluckow
328
21k
Fontdeck: Realign not Redesign
paulrobertlloyd
83
5.4k
I Don’t Have Time: Getting Over the Fear to Launch Your Podcast
jcasabona
32
2.1k
CoffeeScript is Beautiful & I Never Want to Write Plain JavaScript Again
sstephenson
160
15k
Understanding Cognitive Biases in Performance Measurement
bluesmoon
27
1.6k
Building an army of robots
kneath
303
45k
Typedesign – Prime Four
hannesfritz
41
2.5k
Being A Developer After 40
akosma
89
590k
Site-Speed That Sticks
csswizardry
4
410
RailsConf & Balkan Ruby 2019: The Past, Present, and Future of Rails at GitHub
eileencodes
134
33k
Adopting Sorbet at Scale
ufuk
75
9.2k
Raft: Consensus for Rubyists
vanstee
137
6.8k
Transcript
Statistical Thinking for Data Science Chris Fonnesbeck Vanderbilt University
None
None
21/22 falling 7+ stories survived
2 fell together
40% at night
“Even more surprising, the longer the fall, the greater the
chance of survival.”
2 to 32 stories (average = 5.5)
?
"... 132 such victims were admitted to the Animal Medical
Center on 62nd Street in Manhattan ..."
"Found" Data
convenience sample
Missing Data
Representative
Statistical Issues
Big Data
“With enough data, the numbers speak for themselves ” Chris
Anderson, Wired
Alfred Landon
Literary Digest Straw Poll
"Next week, the first answers from these ten million will
begin the incoming tide of marked ballots, to be triple-checked, verified, five-times cross-classified and totalled."
2.4 million returns
41 - 55
None
George Gallup
Sampled 50,000
66%
Random Sampling
None
Bias
None
None
Self-selection Bias
None
For some estimate of unknown quantity ,
p = 0.5 sample_sizes = [10, 100, 1000, 10000, 100000]
replicates = 1000 biases = [] for n in sample_sizes: bias = np.empty(replicates) for i in range(replicates): true_sample = np.random.normal(size=n) negative_values = true_sample<0 missing = np.random.binomial(1, p, n).astype(bool) observed_sample = true_sample[~(negative_values & missing)] bias[i] = observed_sample.mean() biases.append(bias)
None
Accuracy Mean Squared Error
“The numbers have no way of speaking for themselves” Nate
Silver
White House Big Data Partners Workshop
White House Big Data Partners Workshop 19 Participants 0 Statisticians
NSF Working Group on Big Data
NSF Working Group on Big Data 100 experts convened 0
statisticians
Moore Foundation Data Science Environments
Moore Foundation Data Science Environments 0 directors with statistical expertise
NIH BD2K Executive Committee
NIH BD2K Executive Committee 17 committee members 0 statisticians
Feeling left out?
It's our own fault
“Almost everything you learned in your college statistics course was
wrong”
Typical introductory statistics syllabus 1.Descriptive statistics and plotting
Typical introductory statistics syllabus 1.Descriptive statistics and plotting 2.Basic probability
Typical introductory statistics syllabus 1.Descriptive statistics and plotting 2.Basic probability
3.Hypothesis testing
Typical introductory statistics syllabus 1.Descriptive statistics and plotting 2.Basic probability
3.Hypothesis testing 4.Experimental design
Typical introductory statistics syllabus 1.Descriptive statistics and plotting 2.Basic probability
3.Hypothesis testing 4.Experimental design 5.ANOVA
Statistical Hypothesis Testing
None
None
Test Statistic
T-statistic
None
None
None
p-value
None
None
false positive rate
"The value for which , or 1 in 20, is
1.96 or nearly 2; it is convenient to take this point as a limit in judging whether a deviation ought to be considered significant or not." R.A. Fisher
p-value
the probability that the observed differences are due to chance
the probability that the observed differences are due to chance
a measure of the reliability of the result
a measure of the reliability of the result
the probability that the null hypothesis is true
the probability that the null hypothesis is true
"If an experiment were repeated infinitely, p represents the proportion
of values more extreme than the observed value, given that the null hypothesis is true."
H0 : Mean duckling body mass did not differ among
years.
H0 : Mean duckling body mass did not differ among
years.
H0 : The prevalence of autism spectrum disorder for males
and females were equal.
H0 : The prevalence of autism spectrum disorder for males
and females were equal.
H0 : The density of large trees in logged and
unlogged forest stands were equal
H0 : The density of large trees in logged and
unlogged forest stands were equal
Statistical Straw Man
Statistical hypotheses are not interesting
Hypothesis tests are not decision support tools
Multiple Comparisons
None
Family-wise Error Rate >>> 1. - (1. - 0.05) **
20 0.6415140775914581
import seaborn as sb import pandas as pd n =
20 r = 36 df = pd.concat([pd.DataFrame({'y':np.random.normal(size=n), 'x':np.random.random(n), 'replicate':[i]*n}) for i in range(r)]) sb.lmplot('x', 'y', df, col='replicate', col_wrap=6)
None
Statistically Significant!
None
"Despite a large statistical literature for multiple testing corrections, usually
it is impossible to decipher how much data dredging by the reporting authors or other research teams has preceded a reported research finding."
What's the Alternative?
Build models and use them to estimate things we care
about
Effect size estimation
Data-generating Model
None
None
Florida manatee Trichechus manatus
None
None
None
occupied?
occupied? available?
occupied? available? seen?
None
Estimating visibility
None
None
None
None
None
None
None
Bayesian Statistics
None
None
Bayes' Formula
Probabilistic Modeling
Evidence-based Medicine
ASD Interventions Research 19 independent studies 27 different interventions
None
None
None
None
None
None
None
None
None
None
None
“While everyone is looking at the polls and the storm,
Romney’s slipping into the presidency. ”
None
Heirarchical modeling
Pollster effects
None
None
None
None
Data Science
Data
Science
Those who ignore statistics are condemned to re-invent it. --
Brad Efron