statistics: the compass for navigating a data …big data also supplies more raw material for...

Statistics:The Compass for Navigating

a Data-Centric World

Marie Davidian

Department of StatisticsNorth Carolina State University

January 11, 2013

Statistics2013 Video

Available at http://statistics2013.org

http://www.youtube.com/watch?feature=player_embedded&v=nTBZuQR7dRc

http://statistics2013.org

Triumph of the geeks

Nate Silver predicted the outcome of the 2012 US presidentialelection in all 50 states

using . . .

Statisticshttp://fivethirtyeight.blogs.nytimes.com/

Silver used a statistical model to combine the results ofstate-by-state polls, weighting them according their previousaccuracy, and to simulate many elections and estimateprobabilities of the outcome

http://fivethirtyeight.blogs.nytimes.com/


Nate Silver predicted the outcome of the 2012 US presidentialelection in all 50 states using . . .

Statistics





Nate Silver predicted the outcome of the 2012 US presidentialelection in all 50 states using . . .

Statisticshttp://fivethirtyeight.blogs.nytimes.com/




Others did, too. . .

“Dynamic Bayesian forecasting of presidential elections in thestates,” by Drew A. Linzer, Journal of the American StatisticalAssociation, in press


“Nate Silver-led statistics men crush pundits in election”– Bloomberg Businessweek

“Nate Silver has made statistics sexy again”– Associated Press

“Drew Linzer: The stats man who predicted Obama’s win”– BBC News Magazine

“The allure of the statistics field grows”– Boston Globe

But the interest in statistics didn’t start with theUS elections. . .

Statistics in the news

New York Times, August 6, 2009

“I keep saying that the sexy job in the next 10 years will bestatisticians” – Hal Varian, Chief Economist, Google


New York Times, January 26, 2012

“I went to parties and heard a little groan when people heardwhat I did. Now they’re all excited to meet me” – Rob

Tibshirani, Department of Statistics, Stanford University


New York Times, February 11, 2012

“Statistics are interesting and fun. It’s cool now” – AndrewGelman, Department of Statistics, Columbia University


The Wall Street Journal, December 28, 2012

Carl Bialik, The Numbers Guy

Data, data, and more data

Why is there so much talk of statistics andstatisticians?

Data• Administrative (e.g., tax records), government surveys• Genomic, meteorological, air quality, seismic, . . .• Electronic medical records, health care databases• Credit card transactions, point-of-sale, mobile phone• Online search, social networks• Polls, voter registration records

A veritable tsunami/deluge/avalanche of data



Data

• Administrative (e.g., tax records), government surveys• Genomic, meteorological, air quality, seismic, . . .• Electronic medical records, health care databases• Credit card transactions, point-of-sale, mobile phone• Online search, social networks• Polls, voter registration records




Data• Administrative (e.g., tax records), government surveys• Genomic, meteorological, air quality, seismic, . . .• Electronic medical records, health care databases• Credit card transactions, point-of-sale, mobile phone• Online search, social networks• Polls, voter registration records


Demand

2011 McKinsey Global Institute report:

Big data: The next frontier for innovation,competition, and productivity

“A significant constraint. . . will be a shortage of . . . people withdeep expertise in statistics and data mining. . . a talent gap of

140K - 190K positions in 2018 (in the US)”

http://www.mckinsey.com/insights/mgi/research/technology and innovation/big data the next frontier for innovation

Opportunities and challenges

• Our ability to collect, store, access, and manipulate vastand complex data is ever-improving

• The potential benefits to science and society of learningfrom these data are enormous

• However, Big Data does not automatically meanBig Information

• Science, decision-making, and policy formulation requirenot only prediction and finding associations and patterns,but uncovering causal relationships

• Which, as we’ll discuss later, is not so easy. . .

Perils

From “The Age of Big Data”With huge data sets and fine-grained measurement,. . . there isincreased risk of “false discoveries.” The trouble with seeking ameaningful needle in massive haystacks of data, says TrevorHastie, a statistics professor at Stanford, is that “many bits ofstraw look like needles.”

Big Data also supplies more raw material for statisticalshenanigans and biased fact-finding excursions. It offers ahigh-tech twist on an old trick: I know the facts, now let’s find’em. That is, says Rebecca Goldin, a mathematician at GeorgeMason University, “one of the most pernicious uses of data.”

Critical need

Sound, objective methods for modeling,analysis, and interpretation

Statistics

While Big Data have inspired considerable current interest instatistics, statistics has been fundamental in numerous areas ofscience, business, and government for decades

Roadmap

• A brief history

• Statistical stories• Our data-rich future

Roadmap

• A brief history• Statistical stories

• Our data-rich future

Roadmap

• A brief history• Statistical stories• Our data-rich future

What is statistics?

Statistics: The science of learning from dataand of measuring, controlling, andcommunicating uncertainty

The path to what is now the formal discipline of statisticalscience is long and winding. . .

Origins – pre-1700

• Sporadic accounts of measurement and data collectionand interpretation date back as early as 5 B.C.

• But it was not until the the mid-1600s that themathematical notions of probability began to be developedby (mainly) mathematicians and physicists (e.g., BlaisePascal), often inspired by games of chance

• The first formal attempt to summarize and learn from datawas by John Graunt, who created a precursor to modernlife tables used in demography

• Christiaan Huygens was among the first to connect suchdata analysis to probability

Origins – 1700-1750

• From 1700 to 1750, many key results in classicalprobability that underlie statistical theory were derived

• Jakob Bernoulli– law of large numbers, the Bernoulli andbinomial probability distributions

• Abraham de Moivre – The Doctrine of Chances, precursorto the central limit theorem

• Daniel Bernoulli – expected utility, applications ofprobability to measurement problems in astronomy

Milestone events – 1750-1820

• Thomas Bayes’ 1763 An essay towards solving a problemin the Doctrine of Chances presented a special case ofBayes’ theorem (posthumously)

• Arien-Marie Legendre described the method of leastsquares in 1805

●

●

●

●

●

●

●

●

●

●

●

●

●


• Thomas Bayes’ 1763 An essay towards solving a problemin the Doctrine of Chances presented a special case ofBayes’ theorem (pothumously)


●

●

●

●

●

●

●

●

●

●

●

●

●


• Thomas Bayes’ 1763 An essay towards solving a problemin the Doctrine of Chances presented a special case ofBayes’ theorem (pothumously)



• Carl Fredrich Gauss connected least squares to Bayestheorem in 1809

• Pierre-Simon Laplace derived the central limit theorem andconnected the normal probability distribution to leastsquares in 1810

More milestones – 1820-1900

• Aldolphe Quetelet pioneered the statistical analysis ofsocial science data – the “average man” (1835) and thenormal distribution as a model for measurements (1842)

• The Royal Statistical Society (1834) and AmericanStatistical Association (1839) were founded

• Francis Galton introduced regression analysis (1885) andcorrelation (1888)

• Karl Pearson established the field of biometry anddeveloped fundamental methods, and founded the firststatistical journal, Biometrika (1901)

Modern statistics – 1900-1950s

The modern discipline of statistics was reallyestablished only in the twentieth century

• William Gosset (“Student”), a brewer for Guinness inDublin, derived the Student’s t distribution in 1908

• In the 1920s, Ronald Fisher developed many fundamentalconcepts, including the ideas of statistical models andrandomization, theory of experimental design, the methodof analysis of variance, and tests of significance

• In the 1930s, Jerzy Neyman and Egon Pearson developedthe theory of sampling, the competing approach ofhypothesis testing, and the concept of confidence intervals

• Experimental design became a mainstay of agriculturalresearch


The modern discipline of statistics was reallyestablished only in the twentieth century• William Gosset (“Student”), a brewer for Guinness in

Dublin, derived the Student’s t distribution in 1908

• In the 1920s, Ronald Fisher developed many fundamentalconcepts, including the ideas of statistical models andrandomization, theory of experimental design, the methodof analysis of variance, and tests of significance




The modern discipline of statistics was reallyestablished only in the twentieth century• William Gosset (“Student”), a brewer for Guinness in

Dublin, derived the Student’s t distribution in 1908• In the 1920s, Ronald Fisher developed many fundamental

concepts, including the ideas of statistical models andrandomization, theory of experimental design, the methodof analysis of variance, and tests of significance




• Fisher/Neyman-Pearson established the paradigm offrequentist statistical inference that is used today

• Also in the 1930s, Bayesian statistical inference wasdeveloped by Bruno de Finetti and others

• In the 1940s, many departments of statistics wereestablished at universities in the US and Europe

• And fundamental theory of statistical inference waspursued by Wald, Cramer, Rao and many others

Modern statistics to the present

From the 1950s on, there were numerousadvances in theory, methods, and application• The advent of medical statistics and epidemiological

methods (Richard Doll, Austin Bradford Hill)• The development of methods for analysis of censored

time-to-event data (Paul Meier, D.R. Cox)• The use of the theory of sampling to design surveys and

the US census (Jerzy Neyman, Morris Hansen)• The adoption of statistical quality control and experimental

design in industry (W. Edwards Deming, George Box)• Exploratory data analysis (John Tukey)• And many, many more. . .


Computing fundamentally altered the field ofstatistics forever• Complex calculations became feasible• Much larger and more complicated data sets could be

created and analyzed• Sophisticated models and methods could be applied

• Statistical software implementing popular methods becamewidespread (e.g., SAS, developed at NC State in the1960s/70s)

• Simulation to investigate performance of statisticalmethods became possible

• Bayesian statistical methods became feasible in complexsettings (Markov chain Monte Carlo – MCMC)


Computing fundamentally altered the field ofstatistics forever• Complex calculations became feasible• Much larger and more complicated data sets could be

created and analyzed• Sophisticated models and methods could be applied• Statistical software implementing popular methods became

widespread (e.g., SAS, developed at NC State in the1960s/70s)

• Simulation to investigate performance of statisticalmethods became possible

• Bayesian statistical methods became feasible in complexsettings (Markov chain Monte Carlo – MCMC)

Today

Statistical methods are used routinely inscience, industry/business, and government• Pharmaceutical companies employ statisticians, who work

in all stages of drug development

• Statisticians are ubiquitous in medical and public healthresearch, working with health sciences researchers todesign studies, analyze data, and draw conclusions

• Google, Facebook, LinkedIn, credit card companies, globalretailers employ statisticians to develop and implementmethods to mine their vast data

• Government science, regulatory, and statistical agenciesemploy statisticians to design surveys, make forecasts,develop estimates of income, review new drug applications,assess evidence of health effects of pollutants, . . .

Today

Statistical methods are used routinely inscience, industry/business, and government• Pharmaceutical companies employ statisticians, who work

in all stages of drug development• Statisticians are ubiquitous in medical and public health

research, working with health sciences researchers todesign studies, analyze data, and draw conclusions

• Google, Facebook, LinkedIn, credit card companies, globalretailers employ statisticians to develop and implementmethods to mine their vast data

• Government science, regulatory, and statistical agenciesemploy statisticians to design surveys, make forecasts,develop estimates of income, review new drug applications,assess evidence of health effects of pollutants, . . .

Statistical stories

Some diverse examples where statistics andstatisticians are essential. . .

The controlled clinical trial

The gold standard study for comparison oftreatments (a question of cause and effect)

• An experiment designed to compare a new treatment to acontrol treatment

• Subjects are randomized to receive one treatment or theother⇒ unbiased, fair comparison using statisticalmethods (hypothesis testing)

• In addition, blinding, placebo• The first such clinical trial was conducted in the UK by the

Medical Research Council in 1948, comparingstreptomycin+bed rest to bed rest alone in tuberculosis

• In 1954, 800K children in the US were randomized to theSalk polio vaccine or placebo to assess the vaccine’seffectiveness in preventing paralytic polio


The gold standard study for comparison oftreatments (a question of cause and effect)• An experiment designed to compare a new treatment to a

control treatment• Subjects are randomized to receive one treatment or the

other⇒ unbiased, fair comparison using statisticalmethods (hypothesis testing)

• In addition, blinding, placebo• The first such clinical trial was conducted in the UK by the

Medical Research Council in 1948, comparingstreptomycin+bed rest to bed rest alone in tuberculosis

• In 1954, 800K children in the US were randomized to theSalk polio vaccine or placebo to assess the vaccine’seffectiveness in preventing paralytic polio


• In 1969, evidence from a randomized clinical trial becamemandatory for a new product to receive approval from theUS Food and Drug Administration (FDA)

• Because a trial involves only a sample of patients from theentire population, the results are subject to uncertainty

• Statistical methods are critical for determining the samplesize required to ensure that a real difference can bedetected with a specified degree of confidence

• Which is why regulatory bodies like the FDA employ 100sof statisticians

• In the last 4 decades, statisticians have developed newmethods to handle ethical and practical considerations

• E.g., group sequential trials that allow interim analyses atwhich the trial can be stopped early without compromisingthe ability to make a valid comparison

National forest inventory

Next stop, Bhutan• The Kingdom of Bhutan, in South Asia, transitioned to a

constitutional democracy in 2008• The new constitution mandates that Bhutan maintain 60%

forest cover in perpetuity• A National Forest Inventory was called for. . .

• My friend Tim Gregoire of Yale University, an expert inforest biometry, was consulted to help plan and implementBhutan’s comprehensive NFI


Next stop, Bhutan• The Kingdom of Bhutan, in South Asia, transitioned to a

constitutional democracy in 2008• The new constitution mandates that Bhutan maintain 60%

forest cover in perpetuity• A National Forest Inventory was called for. . .• My friend Tim Gregoire of Yale University, an expert in

forest biometry, was consulted to help plan and implementBhutan’s comprehensive NFI


A NFI is an assessment based on statisticalsampling and estimation of the forest resourcesof a nation• Set policy on forest resource management• Monitor biodiversity, habitat type and extent, land

conversion rates• Measure quantity/quality of wood fiber for commodities• Measure non-wood forest products• Measure carbon storage and change• Reference spatially where resources are located

Statistics is critical to developing the sampling plan for bothremote sensing and field data and to estimation of abundanceof resources based on 100s of measurements

Pharmacokinetics

What’s behind a drug label?• A drug should be safe and effective• Labeling provides guidance on dose, conditions under

which a drug should/should not be taken• Partly behind this – pharmacokinetics (PK), the science of

“what the body does to the drug”

• Key: Understanding Absorption, Distribution, Metabolism,Excretion in the population and how these processes varyacross patients and are altered by conditions

• Statistical modeling is an integral part of the science

Pharmacokinetics

What’s behind a drug label?• A drug should be safe and effective• Labeling provides guidance on dose, conditions under

which a drug should/should not be taken• Partly behind this – pharmacokinetics (PK), the science of

“what the body does to the drug”• Key: Understanding Absorption, Distribution, Metabolism,

Excretion in the population and how these processes varyacross patients and are altered by conditions

• Statistical modeling is an integral part of the science

Pharmacokinetics

A hierarchical statistical model that allows these processes tovary across patients and conditions is fitted to drugconcentration-time data

Pharmacokinetics

Conc(t) =ka Dose

V (ka − Cl/V )[exp{−(Cl/V )t} − exp(−kat)]

ka = absorption rate, V = volume of distribution, Cl = clearance

Forensic science

An area where statisticians and better statisticsare desperately needed!• Fingerprints, DNA analysis, bite marks, firearm toolmarks,

hair specimens, writing samples, toxicological analysis,. . .• Laboratory- or expert interpretation-based• 2009 US National Academy of Sciences report

• The report cites examples of lack of sufficient recognitionof sources of variability and their effects on uncertainties inmany types of forensic science analyses. . .

Forensic science

“With the exception of nuclear DNA analysis, however, no forensicmethod has been rigorously shown to have the capacity toconsistently, and with a high degree of certainty, demonstrate aconnection between evidence and a specific individual or source.”

“A body of research is required to establish the limits and measuresof performance and to address the impact of sources of variabilityand potential bias.”

“The development of quantifiable measures of uncertainty in theconclusions of forensic analyses . . . and of quantifiable measures ofthe reliability and accuracy of forensic analyses (are needed).”

Basically, the report recommends that current and new forensicpractices should be developed and assessed using properlydesigned experiments and statistical methods!

The hazards of haphazard data

When data are simply observed and collected,without a principled design and randomization,be wary!• Investigations of causal relationships can be compromised

by confounding• E.g., comparison of the effects of competing treatments• When individual patients and their providers decide which

treatment to take, there may be factors that are associatedwith both the choice of treatment and outcome

• Failure to recognize/identify such confounding factors canlead to misleading conclusions

Simpson’s paradox

Data on 2 treatments from a healthcare database

Ave

rage

Out

com

e

Avg Trt A

Avg Trt B

Simpson’s paradox


Ave

rage

Out

com

e

Trt A

Trt A

Trt B

Trt B

Male Female

Simpson’s paradox


Ave

rage

Out

com

e

Trt A

Trt A

Trt B

Trt B

Avg Trt A

Avg Trt B

A: 80%/20% M/F B: 20%/80% M/F

Male Female

Confounding and other threats

• Statistical methods are available to take confounding intoappropriate account

• . . . but the confounding factors must be recorded in thedatabase!

Other threats• Missing information – why are some factors not recorded

for some individuals?• Drop out – sicker patients may disappear sooner in a

longitudinal study• Etc

Comparative effectiveness research, which strivesto recommend best uses for existing treatment throughanalyses of such databases, requires statistics!

Confronting our data-rich future

I hope I have convinced you that statistics and statisticians areessential to our data-rich future!

Big Data have enormous potential for new generating newknowledge and improving human welfare. However, Big Datawithout statistics have enormous potential to mislead.

“The future demands that scientists, policy-makers, and thepublic be able to interpret increasingly complex information andrecognize both the benefits and pitfalls of statistical analysis.Embedding statistics in science and society will pave the routeto a data-informed future, and statisticians must lead thischarge.”– Davidian and Louis, Science, April 6, 2012

2013 – the International Year of Statistics

A celebration of the contributions of statistics islong overdue!



References and further reading

Aldrich, J. Figures from the history of probability and statistics.http://www.economics.soton.ac.uk/staff/aldrich/Figures.htm

Davidian, M. and Louis, T.A. (2012). Why statistics? Science, 336, 12.

Feinberg, S.E. (1992). A brief history of statistics in three and onehalf chapters: A review essay. Statistical Science, 7, 208–225.

Stigler, S.M. (1986). The History of Statistics: The Measurement ofUncertainty Before 1900. Harvard University Press.

Personalized Medicine:The Right Treatment for the

Right Patient

Marie Davidian

Department of StatisticsNorth Carolina State University

January 11, 2013

“The right treatment for the right patient(at the right time) ”

So why is a statistician talking to you aboutpersonalized medicine?

My goal: To convince you that the quantitativesciences, and especially statistics, are essentialin the quest for personalized medicine!

Roadmap

• Some background

• What is personalized medicine?• Statistics, mathematics, and personalized

medicine

Roadmap

• Some background• What is personalized medicine?

• Statistics, mathematics, and personalizedmedicine

Roadmap

• Some background• What is personalized medicine?• Statistics, mathematics, and personalized

medicine

Modern expectationThere should be a treatment for that!

“Treatments” – drugs, biologic products, medicaldevices, surgical procedures, behavioraltherapies – are omnipresent in today’s world• Cholesterol-lowering medications, anti-platelet therapies• Anti-depressants, anti-psychotics, cognitive therapies• Chemotherapies, tamoxifen (Nolvadex), bevacizumab

(Avastin), rituximab (Retuxan)• Antiretroviral therapies, e.g., PIs, NNRTIs, FIs• Artificial hips, implants

How are treatments developed?

Who decides if they “work?” How is thisdecided?

How are treatments developed and evaluated?

Main players• Pharmaceutical, biotechnology, device companies• University and government research

Who decides? And how? Today• US: Food and Drug Administration (FDA)• EU: European Medicines Agency (EMA)• Japan: Pharmaceuticals and Medical Devices Agency• International Conference on Harmonisation (ICH)• Safety – can people take it?• Efficacy – does it do anything in humans?• Effectiveness – is it better or at least as good as what is

currently available?• Do the benefits outweigh the risks?

Regulatory process

Today, the process of deciding is highlyregulated

But it wasn’t always like this. . .

No regulatory process

At the beginning of the 20th century, there wasessentially no regulation anywhere!• Manufacturers could advertise any product as a treatment

for anything, with no evidence• Drugs like opium, heroin, cocaine were freely available• No requirements for labeling or a list of ingredients• A “free-for-all”

Path to modern regulatory agencies

Little by little, steps were taken• To create what is now the modern FDA in the US (1927)• To require evidence of safety in the US (1938)• To introduce the concept of a prescription in the US (1951)

The bombshell: 1962 – Thalidomide• Anecdotal reports of birth defects in Europe• A FDA medical officer argued for keeping the drug off the

US market⇒ public support for stronger drug regulation• Legislation enacted in the US requiring demonstration of

safety and effectiveness for the first time by “substantialevidence” from “well-controlled studies”

Today

1962 to present• The Declaration of Helsinki was developed by the World

Medical Association to set forth ethical principles forresearch involving human subjects (1964)

• Similarly, the Belmont Report in the US (1979)• The current, highly regulated process of bringing a new

treatment to market was established

Fundamental – the controlled clinical trial• Evaluation of effectiveness• Comparison of a new treatment to standard of care• Comparison of existing treatments to establish new uses


Basics of a “confirmatory” clinical trial• A sample of subjects with the disease/disorder is recruited• 100s to 1000s of subjects• Subjects are randomized to treatments under study⇒ eliminate bias, allow fair comparison

• A clinical outcome is ascertained for each subject• E.g., survival time in cancer, viral load level in human

immunodeficiency virus (HIV) infection after 1 year


Effectiveness• Compare some summary measure of clinical outcomes

between/among treatments• E.g., the average• “Is the average outcome if all patients in the population

took treatment A different from (better than) that if they allinstead took treatment B?”

• Use statistical methods to evaluate the strength of theevidence in the data from the sample supporting a realdifference in the population

Results

Thus, usually• Assesment of effectiveness and regulatory approval are

based on a summary measure (e.g., an average) acrossthe entire population

• Statistics is key

Success• Countless treatments have been approved on this basis• And have benefited numerous patients

However. . .• All patients are not created equal

Patient heterogeneity


We’re all different• Physiological, demographic characteristics• Medical history• Genetic/genomic characteristics

What works for a patient with one set ofcharacteristics might not work for another

An (admittedly contrived) example

Average outcome in the patient population• Larger outcomes are better (survival time)• If all patients took treatment A = 9 months• If all patients took treatment B =18 months• Treatment B is better on average

Genetic mutation• 20% have it, 80% don’t• If all patients took treatment A = (0.2)(25) + (0.8)(5) = 9• If all patients took treatment B = (0.2)(10) + (0.8)(20) = 18• That is, patients with the mutation do much better on

treatment A! (25 months vs. 10 months)



Genetic mutation• 20% have it, 80% don’t

• If all patients took treatment A = (0.2)(25) + (0.8)(5) = 9• If all patients took treatment B = (0.2)(10) + (0.8)(20) = 18• That is, patients with the mutation do much better on




Genetic mutation• 20% have it, 80% don’t• If all patients took treatment A = (0.2)(25) + (0.8)(5) = 9

• If all patients took treatment B = (0.2)(10) + (0.8)(20) = 18• That is, patients with the mutation do much better on




Genetic mutation• 20% have it, 80% don’t• If all patients took treatment A = (0.2)(25) + (0.8)(5) = 9• If all patients took treatment B = (0.2)(10) + (0.8)(20) = 18

• That is, patients with the mutation do much better ontreatment A! (25 months vs. 10 months)



Genetic mutation• 20% have it, 80% don’t• If all patients took treatment A = (0.2)(25) + (0.8)(5) = 9• If all patients took treatment B = (0.2)(10) + (0.8)(20) = 18• That is, patients with the mutation do much better on



Moral• “One size does not fit all”• Use a patient’s characteristics to determine which

treatment option might be best for him/her

• Genomic information may hold great potential• Personalized medicine


Moral• “One size does not fit all”• Use a patient’s characteristics to determine which

treatment option might be best for him/her• Genomic information may hold great potential• Personalized medicine


Not so fast . . .• How do we do this?• What are the challenges and possible pitfalls?

Popular persepective on personalized medicine

Subgroup identification and targeted treatment• Can we determine subgroups of patients who share

certain characteristics and who are more likely to do betteron a particular treatment than on others?

• Can biomarkers be developed to identify such patients?• In fact, can a new treatment be developed to target a

subgroup that is likely to benefit?• Can clinical trials and approval be focused on particular

subgroups of patients?

Focus on finding and treating a subgroup

Popular persepective on personalized medicine

Another perspective on personalized medicine

Can we determine how best to treat the entirepopulation?• Given information on any patient’s characteristics, can we

determine the treatment most likely to benefit him/her?• In fact, can we come up with “rules” that take a patient’s

chacteristics as input and output the best option forhim/her?

Focus on treating everyone

Challenge

In either case• It’s all about identifying “tailoring variables”• Knowledge of the biology integrated with statistics

Finding tailoring variables

Ave

rage

Out

com

e

Trt A

Trt A

No Mutation Mutation


Ave

rage

Out

com

e

Trt A

Trt A

Trt B

Trt B



Ave

rage

Out

com

e

Trt A

Trt A



Ave

rage

Out

com

e

Trt A

Trt A

Trt B

Trt B


Result

Moral• Need to identify tailoring variables• This is a statistical problem. . .

Challenge, more precisely

High dimensional data!• Must sift through 1000s of characteristics to identify the

right combination of key tailoring variables• Based on data from a sample of patients (10s, 100s,

1000s)

Challenge, more precisely

Pitfalls• Computational complexity• Chance to miss important characteristics• Chance of false discovery• Statistical methods must be used

Can we treat everyone “optimally?”

Clinical practice: Treatment decisions over time• Fixed schedule• Event(s) necessitating a decision

Clinical decision-making• Clinical judgment used to synthesize all information

available, make a “personalized” treatment decision• Can this be formalized?• That is, can we construct decision rules?

Cancer treatment

Two decision points• Decision 1: Induction chemotherapy (C)• Decision 2: Maintenance treatment (M) for patients who

respond, Salvage chemotherapy (S) for those who don’t• Several options for each• Goal: Prolong survival

Sequential decision-making

• Decision rule 1: Genetic/genomic profile, demographics,physiological characteristics, medical history,. . .⇒ which of2 chemotherapies C to use

• Decision rule 2: Previous info + responder status,intermediate physiological/clinical measures, sideeffects,. . .⇒ which of 2 maintenance therapies M(responders) or 2 salvage chemotherapies S(non-responders) to use

Decision rules

• Decision rule 1: “If age < 50, progesterone receptor level< 10 fmol, RAD51 mutation, then give C1, else, give C2”

• Decision rule 2: “If patient responds, age < 60, CEA > 10ng/mL, progesterone receptor level < 8 fmol, give M1, else,give M2; if does not respond, age > 65, P53 mutation, CA15-3 > 25 units/mL, then give S1, else, give S2”

Statistical problem

Construct (estimate) rules from data• At each decision, identify the tailoring variables and the

right function of them to give a decision rule• Goal: Find the decision rules that would lead to best

expected outcome if followed by all patients• Each rule must take account of what might happen later

when deciding what to do now• How to do this: Statistical modeling, dynamic programming

Treatment of acute HIV infection

47 year old male goes to the ER• 102.5 ◦F fever, headache, nausea/vomiting, rash, . . .• MSM, recent unprotected sex, . . .• Tests for cytomegalovirus (CMV), Epstein-Barr virus

(EBV), influenza negative• HIV test positive• HIV RNA (viral load ) > 750,000 copies/ml• CD4+ T cell count = 432 cells/µL

Diagnosis – Acute HIV infection


Should this patient be started on antiretroviraltherapy (ART)?

• Disadvantages: Cost, side effects, eventual drugresistance, limit future ART options

• Advantages: “Train” the immune system through cycles oftreatment “interruption” – cycles of treatment and viralexposure may allow patient to maintain control of virus

More generally• Can we determine the “best” way to use ART to manage

the infection and prolong time to AIDS?


Should this patient be started on antiretroviraltherapy (ART)?• Disadvantages: Cost, side effects, eventual drug

resistance, limit future ART options• Advantages: “Train” the immune system through cycles of

treatment “interruption” – cycles of treatment and viralexposure may allow patient to maintain control of virus

More generally• Can we determine the “best” way to use ART to manage

the infection and prolong time to AIDS?

Mathematical modeling

HIV dynamic model• Represent mechanisms involved in virus-immune system

interaction mathematically• System of differential equations• Over-simplification of complex biology, but can be useful• Predict viral load, CD4 count at any time under any ART

strategy

Dynamical system model

Dynamical system model

T1 = λ1 − d1T1 − {1− ε1u(t)}k1VIT1

T ∗1 = {1− ε1u(t)}k1VIT1 − δT ∗

1 −m2ET ∗1

T2 = λ2 − d2T2 − {1− f ε1u(t)}k2VIT2

T ∗2 = {1− f ε1u(t)}k2VIT2 − δT ∗

2 −m2ET ∗2

VI = {1− ε2u(t)}103NT δ(T ∗1 + T ∗

2 )− cVI

−{1− ε1u(t)}ρ1103k1T1VI

− {1− f ε1u(t)}ρ2103k2T2VI

VNI = ε2u(t)103NT δ(T ∗1 + T ∗

2 )− cVNI

E = λE +bE(T ∗

1 + T ∗2 )

(T ∗1 + T ∗

2 ) + KbE −

dE(T ∗1 + T ∗

2 )

(T ∗1 + T ∗

2 ) + KdE − δEE

Model-based treatment

Add statistics• Fit the model to data on many subjects using statistical

methods• Use the fitted model + control theory to design ART

interruption strategies• Use the model to study the consequences of different ART

interruption strategies on “virtual patients” (simulation)• Study the promising ones on real patients in a clinical trial

Model-based treatment

Wrap-up

• The goal of truly personalized medicine isstill elusive

• But it is attainable!• Combining quantitative sciences (statistics,

mathematics, computer science,. . . ) withbiological, biomedical sciences is one keythat will pave the way

2013 – the International Year of Statistics



statistics: the compass for navigating a data …big data also supplies more raw material for...

Documents