www.chrisbilder.com1 of 36 turning data into knowledge to solve real world problems christopher r....
TRANSCRIPT
www.chrisbilder.com 1 of 36
Turning data into knowledge to Turning data into knowledge to solve real world problemssolve real world problems
Christopher R. Bilder, Ph.D.Department of Statistics
University of Nebraska-Lincolnwww.chrisbilder.com
www.chrisbilder.com 2 of 36
11 years ago…11 years ago… The year is 1993 Pearl Jam records second CD, Vs.,
– Daughter, Go, Elderly Woman Behind the Counter in a Small Town
– Almost 1 million CDs are sold in the first week
Bill Clinton was inaugurated as the 41st president Movies
– Jurassic Park
– Schindler's List
– Sleepless in Seattle
Husker football – Began 1993 by losing the Orange Bowl badly (again)
– 1993 season went undefeated
Math 4750 - Introduction to Probability and Statistics II – 4-5:15PM Tuesdays and Thursdays in DC 164
– Dr. Stephens
www.chrisbilder.com 3 of 36
11 years ago…11 years ago… Actuarial Science!
– Planned to be an actuary when I started college– Internship at National Indemnity Company at 32nd and Harney – Passed 4 exams under old system
Wanted to go on to graduate school– Math?– Actuarial Science?
Hypothesis testing in Math 4750– Use for decision making!– Scientifically prove a hypothesis or statement – Go to graduate school for statistics!
1994 received BS in Mathematics with pre-actuarial science minor from UNO
www.chrisbilder.com 4 of 36
After UNOAfter UNO Went on to graduate school for statistics
– MS 1996 from Kansas State University– PhD 2000 from Kansas State University– Internships at INEEL in Idaho and pharmaceutical company in Kansas
City– Consult with students and professors in
• Institute of Social and Behavorial Research• College of Agriculture
– Taught courses like Statistical Methods I and II (STAT 3000 and 3010)
Assistant Professor at Oklahoma State University– Department of Statistics – 2000-2003
Assistant Professor at UNL– NEW Department of Statistics– 2003-present
www.chrisbilder.com 5 of 36
PurposePurpose Tell you a little about statistics
– Statistics is mainly a graduate discipline– Most statisticians have undergraduate degrees in math
Turning data into knowledge to solve real world problems– 3 actual examples that come from my teaching and research
About statistics at UNL Website (www.chrisbilder.com/statistics) for more information
www.chrisbilder.com 6 of 36
Undergraduate teaching example for a course like STAT 3000 How could you determine which grocery store, Super Wal-
Mart or Albertson’s, has lower average prices?
– Paired or dependent two sample hypothesis test for Wal-Mart - Albertsons
– Sample the same items at each store
Grocery store pricesGrocery store prices
www.chrisbilder.com 7 of 36
Undergraduate teaching example for a course like STAT 3000 How could you determine which grocery store, Dillon’s or
Food-4-Less in Manhattan, KS, has lower average prices?
– Paired or dependent two sample hypothesis test for Dillon’s - Food-4-Less
– Sample the same items at each store
Only cereals from Fall 1998– Possible problems described later
Grocery store pricesGrocery store prices
www.chrisbilder.com 8 of 36
Grocery store pricesGrocery store prices Sample:
Item Dillon's Food-4-Less Difference1 Malt-o-meal - Tootie Fruities, 15oz $1.99 $1.84 $0.152 Malt-o-meal - Golden Puffs, 18oz $1.99 $1.84 $0.153 Quaker Oats - Life Cereal: Original, 21oz $3.69 $3.49 $0.204 Cheerios, 20oz $4.59 $4.24 $0.355 Cheerios, 15oz $3.79 $3.50 $0.296 Wheaties, 18oz $3.89 $3.60 $0.297 Kellogg’s Funpack, 8 9/16oz $2.89 $2.67 $0.228 Kellogg’s Variety Pack 9 5/8oz. $3.49 $3.14 $0.359 Kellogg’s Frosted Mini-Wheats Bite Size
19oz$3.49 $2.50 $0.99
10 Kellogg’s Frosted Mini-Wheats, 16oz $2.50 $2.73 -$0.2311 Kellogg’s Frosted Flakes, 15oz $3.19 $2.92 $0.2712 Our Family Frosted Flakes, 20oz. $2.50 $1.90 $0.6013 Kellogg’s Crispix, 12oz. $3.49 $3.20 $0.2914 Our Family - Raisin Bran, 20oz $2.50 $1.92 $0.5815 Kellogg’s Smart Start, 13.3oz $3.49 $3.24 $0.2516 Grape Nuts, 24oz $3.00 $2.85 $0.1517 Frosted Alpha Bits, 15oz $3.00 $2.87 $0.13
www.chrisbilder.com 9 of 36
Grocery store pricesGrocery store prices Do you think there are
mean differences?Item Dillon's Food-4-Less Difference
1 Malt-o-meal - Tootie Fruities, 15oz $1.99 $1.84 $0.152 Malt-o-meal - Golden Puffs, 18oz $1.99 $1.84 $0.153 Quaker Oats - Life Cereal: Original, 21oz $3.69 $3.49 $0.204 Cheerios, 20oz $4.59 $4.24 $0.355 Cheerios, 15oz $3.79 $3.50 $0.296 Wheaties, 18oz $3.89 $3.60 $0.297 Kellogg’s Funpack, 8 9/16oz $2.89 $2.67 $0.228 Kellogg’s Variety Pack 9 5/8oz. $3.49 $3.14 $0.359 Kellogg’s Frosted Mini-Wheats Bite Size
19oz$3.49 $2.50 $0.99
10 Kellogg’s Frosted Mini-Wheats, 16oz $2.50 $2.73 -$0.2311 Kellogg’s Frosted Flakes, 15oz $3.19 $2.92 $0.2712 Our Family Frosted Flakes, 20oz. $2.50 $1.90 $0.6013 Kellogg’s Crispix, 12oz. $3.49 $3.20 $0.2914 Our Family - Raisin Bran, 20oz $2.50 $1.92 $0.5815 Kellogg’s Smart Start, 13.3oz $3.49 $3.24 $0.2516 Grape Nuts, 24oz $3.00 $2.85 $0.1517 Frosted Alpha Bits, 15oz $3.00 $2.87 $0.13
25%
75%
50%
www.chrisbilder.com 10 of 36
Grocery store pricesGrocery store prices Paired two sample hypothesis test
– Ho:Dillon’s - Food-4-Less=0Ha:Dillon’s - Food-4-Less0
– t = 4.77, p-value = 0.0002, 95% C.I.: 0.1644 < Dillon’s - Food-4-Less < 0.4274
– Reject equal mean prices
If price was the only consideration, what store should one shop at?
Assumptions– Prices and selection at these two stores are indicative of all stores– Normal populations– The sample was taken in 1998; what about now?– Finite populations
www.chrisbilder.com 11 of 36
MS report – applying statistics or investigating new methodology– 120 page book!– Reduced version published in Chance in 1998
Find a model to estimate the probability of successfor placekicks in the NFL
Video– January 7, 1996– Playoff game– Indianapolis Colts 10
Kansas City Chiefs 7– Lin Elliott of KC will attempt
a 42 yard field goal to tie the game and send it into overtime
PlacekickingPlacekicking
www.chrisbilder.com 12 of 36
PlacekickingPlacekicking What factors affect the probability of success for NFL
placekicks?– Distance– Pressure – How do you quantitatively measure?– Wind– Grass vs. artificial turf– Dome vs. outdoor stadium
Collected data >1,700 placekicks during the 1995 NFL season
Find the best logistic regression model of the form
where p is the probability of success xi for i=1,…,k are independent variables
i measures the effect of xi on p for i=1,…,k
0 1 1 2 2 k k
0 1 1 2 2 k k
x x x
x x x
ep=
1 e
www.chrisbilder.com 13 of 36
PlacekickingPlacekicking The i’s are parameters which are estimated through
maximum likelihood estimation Estimated model
– Change: lead change = 1, non-lead change = 0– Distance: distance in yards– PAT: point after touchdown = 1, field goal = 0– Wind: windy (speed > 15 MPH) = 1, non-windy = 0
What is the estimated probability of success for Elliott’s field goal?– Conditions:– Estimated probability of success: – 90% confidence interval for probability of success:
0.6298 < p < 0.7402
Change Distance PAT Wind1 42 0 0
p 0.6850
4.4984 0.3306change 0.0807distance 1.2592PAT 2.8778wind 0.0907distance wind
4.4984 0.3306change 0.0807distance 1.2592PAT 2.8778wind 0.0907distance wind
ep=
1 e
www.chrisbilder.com 14 of 36
Estimated probability of success for a field goal (PAT=0)
20 30 40 50 60
0.0
0.2
0.4
0.6
0.8
1.0
Distance in Yards
Est
ima
ted
Pro
ba
bili
ty o
f S
ucc
ess
Estimated probability of success of a field goal (PAT=0)
Change=0, Wind=0Change=1, Wind=0Change=0, Wind=1Change=1, Wind=1
42
0.685
www.chrisbilder.com 15 of 36
Estimated probability of success for a field goal (PAT=0)
20 30 40 50 60
0.0
0.2
0.4
0.6
0.8
1.0
Distance in Yards
Est
ima
ted
Pro
ba
bili
ty o
f S
ucc
ess
Estimated probability of success for a field goal (PAT=0)
Lowest Number of Risk Factors
Estimated Probability90% Confidence Interval
Highest Number of Risk Factors
Estimated Probability90% Confidence Interval
www.chrisbilder.com 16 of 36
PlacekickingPlacekicking UNL Department of Statistics developing statistics in sports
specialty – Dr. David Marx
• Works with the UNL athletic department• January 10, 2004 Omaha World Herald article about his work the
men’s basketball team (available at www.chrisbilder.com/statistics)• His students this semester have worked with NASCAR, Lincoln SE
women’s high school soccer team, and Tendu, Inc. (baseball software company).
– Myself• Placekicking• Modeling 64-team NCAA tournaments
www.chrisbilder.com 17 of 36
HCV prevalenceHCV prevalence MS/PhD research – forwarding statistical theory and methodology Hepatitis C (HCV)
– Viral infection that causes cirrhosis and cancer of the liver
– Since HCV is transmitted through contact with infectious blood, screening blood donors is important to prevent further transmission
Questions:– How can blood be screened in a cost effective and timely manner?
– What proportion of people is inflicted with HCV in a population?
Individual testing– Each blood sample is tested individually
– Problems:
• Costly
• Time
+ or - + or - + or - + or - + or -+ or -
www.chrisbilder.com 18 of 36
Group testing– Pool the blood samples together to form n groups of size s
– If the GROUP sample is negative, then all s people do not have the disease
– If the GROUP sample is positive, then at least ONE of the s people have the disease
• May want to determine who in the group has the disease– Strategy works well when prevalence of a disease is small– Dorfman (1943) – first used to test members of the military for disease
HCV prevalenceHCV prevalence
+ or - + or - + or -
Group 1 Group 2 Group n
www.chrisbilder.com 19 of 36
HCV prevalenceHCV prevalence Notation
– Let p = probability an INDIVIDUAL is HCV positive– Let = probability a GROUP is HCV positive– Let s = group size– Let n = number of groups– Let T be a random variable denoting the number of positive GROUPS
• T has a binomial distribution with “n trials” and “ as the probability of success”
•
– Let Y be an UNOBSERVABLE random variable denoting the number of positive INDIVIDUALS in a group
• Y has a binomial distribution with “s trials” and “p as the probability of success”
•
t n tnf(t) (1 ) for t=0,1,2,...,n
t
s s ysg(y) p (1 p) for y=0,1,2,...,s
y
www.chrisbilder.com 20 of 36
HCV prevalenceHCV prevalence How can we estimate p?
– We observe information about the groups, not individuals! – Maximum likelihood estimate of is = # positive / # of groups– = P(group is positive)
= P(at least one individual is positive)
= 1 – P(no individuals are positive) using complement rule
= 1 – (1-p)s since p = P(individual is positive) and s individuals per group
– Solve for p, p = 1- (1- )1/s
– Use invariance property of maximum likelihood estimates to find
• For fixed sample size, is positively biased• It is unbiased as n
ˆ T /n
1/ s 1/ sMLE
ˆp 1 (1 ) 1 (1 T /n) p
www.chrisbilder.com 21 of 36
HCV prevalenceHCV prevalence Can we find a better estimator?
– Yes
How do we measure “better”?– Let be an estimator of – Bias =
• Which would you prefer for a bias: small or large• Given two estimators, which would you prefer
– Estimator with smaller bias– Estimator with larger bias
– Can compare to competing estimators through the “relative bias”• Let and be two estimators of
•
• If RB > 1, then and would be “better”
ˆE( )
1
2
ˆBias( )RB =
ˆBias( )
1 2
1 2ˆ ˆBias( ) Bias( ) 2
www.chrisbilder.com 22 of 36
HCV prevalenceHCV prevalence New estimators
– Proposed in Tebbs, Bilder, and Moser (Communications in Statistics, 2003) and Bilder and Tebbs (under review in Biometrical Journal)
– Derived through “empirical Bayesian methods”
–
where
and is found from maximizing
–
1/ s
EB2t 1
p 1 1ˆn / s 1
EB1
ˆ ˆ(n / s 1) (n t / s 1/ s)p 1
ˆ ˆ(n t / s) (n / s 1 1/ s)
1 x
0
( ) x e dx
T(n 1) (n t / s)
f (t | )s (n t 1) (n / s)
www.chrisbilder.com 23 of 36
0.00 0.02 0.04 0.06 0.08 0.10
05
1015
n=30 and s=10
p
Rel
ativ
e bi
as
EB1EB2EB3
0.00 0.02 0.04 0.06 0.08 0.10
0.6
0.7
0.8
0.9
1.0
1.1
1.2
n=30 and s=10
p
Rel
ativ
e ef
ficie
ncy
EB1EB2EB3
0.00 0.02 0.04 0.06 0.08 0.10
05
1015
n=80 and s=25
p
Rel
ativ
e bi
as
EB1EB2EB3
0.00 0.02 0.04 0.06 0.08 0.10
0.6
0.7
0.8
0.9
1.0
1.1
1.2
n=80 and s=25
p
Rel
ativ
e ef
ficie
ncy
EB1EB2EB3
0.00 0.02 0.04 0.06 0.08 0.10
05
1015
n=30 and s=10
p
Rel
ativ
e bi
as
EB1EB2EB3
0.00 0.02 0.04 0.06 0.08 0.10
0.6
0.7
0.8
0.9
1.0
1.1
1.2
n=30 and s=10
p
Rel
ativ
e ef
ficie
ncy
EB1EB2EB3
0.00 0.02 0.04 0.06 0.08 0.10
05
1015
n=80 and s=25
p
Rel
ativ
e bi
as
EB1EB2EB3
0.00 0.02 0.04 0.06 0.08 0.10
0.6
0.7
0.8
0.9
1.0
1.1
1.2
n=80 and s=25
p
Rel
ativ
e ef
ficie
ncy
EB1EB2EB3
MLE EB,iˆ ˆRB Bias(p ) /Bias(p ) ; when RB > 1, is better thanEB,ip MLEp
1n=30, s=10
n=80, s=25
1
www.chrisbilder.com 24 of 36
HCV prevalenceHCV prevalence Estimation of HCV prevalence in Xuzhou City, China.
– Data from Liu et al. (Transfusion, 1997)– 1,875 blood donors screened for HCV at the Blood Transfusion Service
in Xuzhou City, China• There were 42 positive blood donors found
– In order to test the usefulness of group testing, blood samples were also pooled
• n = 375 groups• s = 5 individuals per group• t = 37 positive groups
– Point estimates of p, the individual probability of being HCV positive• Using individual data: 42/1875 = 0.0224• Using group data:
EB1p 0.020557MLEp 0.020562
EB2p 0.020534
www.chrisbilder.com 25 of 36
HCV prevalenceHCV prevalence Multiple vector transfer designs
– Swallow (Phytopathology, 1985)– Want to estimate the probability a insect vector transfers a pathogen
(virus, bacteria, etc.) to a plant
Brown planthopper
Whitebacked planthopper
www.chrisbilder.com 26 of 36
HCV prevalenceHCV prevalence Multiple vector transfer designs (continued)
– s insect vectors are transferred to a healthy plant– The plant is the “group”– Observe number of plants which contract the pathogen
Greenhouse
Enclosed test plant
Does not transmit virus
Transmits virus
y=0
y=1y=0
y=1y=0
y=0Planthoppery =0 if plant is negative, 0 if plant is positive
www.chrisbilder.com 27 of 36
HCV prevalenceHCV prevalence New research
– Include independent variables to help model p in a logistic regression model,
– Problem: Do not have the individual outcomes– Help to decide who to retest if get a positive group– Multiple traits
• HCV• HIV• Other disease• Simultaneously model
0 1 1 2 2 k k
0 1 1 2 2 k k
x x x
x x x
ep=
1 e
www.chrisbilder.com 28 of 36
Why statistics?Why statistics? Statistics is used in many diverse areas!
– Statistics is the “science of science”– Florence Nightingale quote:
the most important science in the whole world: for upon it depends the practical application of every other science and of every art: the one science essential to all political and social administration, all education, all organization based on experience, for it only gives results of our experience.
I hope you have an interest to take more statistics courses – UNO– Graduate school in statistics or non-statistics programs
Of course, I want you to consider coming to UNL!
www.chrisbilder.com 29 of 36
Statistics at UNLStatistics at UNL Facts
– July 1, 2003 formed– 11 faculty + 2 more in 2004– No undergraduate major– 40+ graduate students (most MS)– Strong commitment from administration– Hardin Hall on East Campus
www.chrisbilder.com 30 of 36
Statistics at UNLStatistics at UNL
33rd s
t.
Department of Statistics
www.chrisbilder.com 31 of 36
Statistics at UNLStatistics at UNL Background of new students
– A few statistics courses – like UNO MATH 4740 and 4750– Statistics is mainly a graduate discipline– Majority have math degrees
Recommendation for UNO classes– Math 4740 and 4750 Intro. to Probability and Statistics I and II – Math 3300 Numerical Methods – Math 4760 Topics in Modeling– Math 4050 Linear Algebra– Math 4230 and 4240 Mathematical Analysis I and II
• Helpful if you plan to go on for a PhD– Stat 3000 and 3010 Statistical Methods I and II
www.chrisbilder.com 32 of 36
Statistics at UNLStatistics at UNL Recommendation for UNO classes (continued)
– Business administration course: 3140 Business Statistical Applications – Computer science programming courses– Information Systems & Quantative Analysis Department courses
• 4150 Advanced Statistical Methods for IS&T• 8160 Applied Distribution Free Statistics • 8340 Applied Regression Analysis• 9120 Applied Experimental Design and Analysis• 9130 Applied Multivariate Analysis
www.chrisbilder.com 33 of 36
Statistics at UNLStatistics at UNL Assistantships
– Work 16-20 hours a week– Teaching - $13K per school year + tuition (MS students)– Project Fulcrum grants - $30K per school year!
• 6 statistics students over the past 3 years have received grant– Research - variable depending on grants
• Statistics and non-statistics faculty grants
What makes us unique?– Consulting course and help desk– STAT 971 – Statistical Modeling– Statistics in sports and work with UNL athletic department– Consulting - All departments in the College of Agriculture and Natural
Resources – Gallup– Bioinformatics
www.chrisbilder.com 34 of 36
Statistics at UNLStatistics at UNL Where do statistics graduates work?
– Pharmaceutical – Pfizer, Merck – Marketing – Target, Hallmark– Government research labs – INEEL, Los Alamos, Sandia, Argonne– Agriculture - Pioneer Hi-Bred – Consulting firms – Quintiles– Everyone that I have known has had a job offer before they graduated!
Salaries– Non-academic starting (2003 American Statistical Association survey)
– Survey response rate was 23.5% by organizations surveyed– See salary surveys at the American Statistical Association’s website
Degree Sample size 25th 50th 75th
MS 102 45.5K 50K 59KPhD 99 60K 65K 75K
Percentile
www.chrisbilder.com 35 of 36
Statistics at UNLStatistics at UNL Applying for graduate school in statistics
– Send out applications before end of fall semester– Apply to more than one school – Visit schools in fall or early spring– Assistantship offers usually first go out in March
7th Annual UNL Regional Workshop in Mathematical Sciences– Statistics, Mathematics, and Computer Science departments– November 2004– Friday afternoon & evening and Saturday morning– Speakers introducing statistics and jobs in statistics– FUNDING available!
www.chrisbilder.com 36 of 36
Statistics at UNLStatistics at UNL For more information…
– E-mail me at [email protected] or [email protected] • Advice• Sit in on a class
– Website: www.chrisbilder.com/statistics • This PowerPoint presentation• Links to
– Introductory information about being a statistician– Jobs (including internships)– Salary information– List of all Departments of Statistics– Professional societies – MS and PhD course websites that myself and others teach– Newspaper and magazine articles about statistical applications