Association Studies: Statistical And Study Design Issues

1
Lyle J. Palmer, PhD
From Genome to DiseaseNHLBI SymposiumJuly 23/24
2003Washington
Association Studies Statistical And Study Design
Issues
2
This lecture

• Association in context
• Genetic association analysis some statistical
  issues
• Study design Problems and issues
• The future for association analysis

3
Some challenges

• Field is young and changing rapidly
• Literature can be difficult
• Example Haseman Elston is a standard linkage
  method
• 4-5 variations on HE recently proposed
• Hard to compare approaches
• Software often free, but sometimes not well-tested

4
Some challenges - continued

• Methods are sometimes oversold
• Sure thing of the day
• Collecting X affected sib pairs
• Collecting X unrelated cases and controls
• Isolated populations
• SNPs

5
(No Transcript)
6
Broad Genetic Epidemiology Study Design
Categories

• Linkage Analysis
• Follows meiotic events through families for
  co-segregation of disease and particular genetic
  variants
• Large Families
• Sibling Pairs (or other family pairs)
• Works VERY well for Mendelian diseases
• Association Studies
• Detect association between genetic variants and
  disease across families exploits linkage
  disequilibrium
• Case-Control designs
• Cohort designs
• Parents affected child trios (TDT)
• May be more appropriate for complex diseases

7
Allelic architecture and mapping strategy
Unlikely to exist
Magnitude of effect
Fn. Studies
Frequency in population
Slide thanks to D. Altshuler
8
What determines the allelic architecture?
Evolutionary selection
Slide thanks to D. Altshuler
9
Power comparision sib-pair linkage versus
population association
Risch, Nature 2000
10
Association Study Approaches

• Candidate gene search
• Limited variants or haplotypes based on prior
  knowledge expert opinion, linkage peaks

• Genome-wide scan
• Dense set of markers throughout genome

Family-based and population-based designs
11
(No Transcript)
12
Phenotype-genotype association

• In practical terms, an observed statistical
  association between an allele and a phenotypic
  trait will be due to one of three situations
• The finding could be due to chance or artifact,
  e.g., confounding or selection bias
• The allele is in linkage disequilibrium with an
  allele at another locus that directly affects the
  expression of the phenotype or
• The allele itself is functional and directly
  affects the expression of the phenotype.

13
Candidate Allele Testing

• Test markers for association with disease
  predisposition
• One approach perform standard single-locus
  chi-squared tests

By Alleles OR ad/bc Test ?21
By Genotype Test c22
14
Single-marker logistic regression

• H0 bi 0
• Genetic model interpretations
• Assume 11 genotype coding represents genotype
  with lowest absolute risk (baseline)
• b1 b2 0 no association with that
  polymorphism
• b1 0, b2 gt 0 (completely) recessive
• b1 b2 gt 0 (completely) dominant
• 0 lt b1 lt b2 additive or multiplicative
• Note This can be extended through GLM to many
  types of outcomes (rather than simply odds of
  disease/not disease, as above)

15
LD Gene Mapping

• General idea
• Exploit the phenomenon of linkage disequilibrium
  (LD) between alleles of closely linked markers to
  identify genetic regions associated with disease
  status.
• i.e., Test for LD between marker loci and
  disease allele
• LD strength (magnitude) ? 1 / r
• LD will be highest at areas of the genome that
  are closest to the disease locus
• Use this to pinpoint (localize, fine-map) the
  disease gene region
• E.g.
• Fine-mapping
• From linkage analysis, may have 10 cM candidate
  region. Next add dense set of markers in
  significant region and perform LD analysis to
  narrow region much further.
• What about whole-genome approach?
• Some have suggested genome-wide LD studies are
  feasible with densely spaced markers.

16
(No Transcript)
17
SNP association with disease allele
GENE
marker SNP
marker SNP
marker SNP
disease allele
marker SNP

• How closely must SNPs be spaced?
• 30,000 to 1,000,000 SNPs to span the genome?
• Numbers will depend on local haplotype structure,
  amount of LD
• -30kb blocks 100,000 independent
  SNPs to span genome

18
Multiple testing

• Multiple comparisons - 50,000 genes. Even if only
  one functional locus/gene tested, very high
  number of false s
• Solutions
• Simulation Empirical p-values
• Replication

19
What is a haplotype?

• Some definitions
• Haplotype Set of particular alleles at separate
  loci on the same transmitted chromosome

• Linkage Disequilibrium (LD) Association between
  those particular alleles due to their proximity
  on the same chromosome (due to linkage)

• Haplotype-based analyses provide increased
  informativity
• Each allele (or mutation) is associated with a
  particular evolutionary history and will thus
  have a unique chromosomal background, or
  haplotype.
• More Powerful

20
Motivation for Haplotype-based analysis

• Advantage of combinatorial approach
• Haplotypes important from population genetics
  standpoint
• Increase ability to identify regions that are IBD
• Biologically, combinations of alleles in a region
  may be functionally important, so set of variants
  on a chromosome may be the causative composite
  allele rather than a particular nucleotide at a
  particular SNP
• Haplotype analyses can be more powerful than
  single-locus analyses when LD is exploited

21
Haplotype vs Single-locus Analyses

• Consider a 2-locus system with a disease-bearing
  haplotype
• A-dx-B

ORA-B 2.0
Slide thanks to Dani Fallin
22
Haplotype Determination Options

• Collect and genotype family members
• ?150 effort, cost
• Family members not available

• Laboratory-based techniques
• Chromosome isolation
• Long-range PCR
• Limited results
• Time consuming
• Cost-prohibitive
• Statistical estimation
• Sequential rules (Clark, 1990)
• Likelihood-based E-M algorithms
• (Hill, 1974, Long et al, '95, Hawley Kidd '95,
  Excoffier Slatkin '95 Fallin Schork, 2000)

23
Relative importance of low risk alleles

• Population attributable fraction the proportion
  of disease that
• would be eliminated if the allele was eliminated
  from the population
• For GRRlt2, alleles with frequency lt.15 have very
  little impact on disease in the population.

24
Study design

• Targeting SNPs likely to be important in the
  population
• For GRRlt2, alleles/haplotypes with frequency
  lt.10-.15 have little impact on disease in the
  population
• If the goal is to develop predictive or
  diagnostic tests, such alleles are of little
  commercial interest
• Studies can be designed to have high power for
  moderate (? 0.10) allele frequencies and GRR ? 2
  sample sizes on the order of 1000 cases and 1000
  controls are a good start

25
Association Studies Potential Causes of
Inconsistent Results

• Population stratification Differences between
  cases and controls most often cited reason
• Genetic heterogeneity Different genetic
  mechanisms in different populations
• Random error False positive/false negative
  results
• Study design/analysis problems
• Poorly defined phenotypes
• Failure to correct for sub-group analysis and
  multiple comparisons
• Poor control group selection
• Small sample sizes
• Failure to attempt replication
• Silverman Palmer Am J Respir Cell Mol Biol 2000
• Cardon Bell, Nat. Rev. Gen. 2001

26
Population stratification

• If cases and controls have different genetic
  backgrounds (are from different genetic
  sub-populations),
• There may be inherent gene frequency differences,
  increasing the possibility of a false positive
  (or negative) result

27
Knowler, W. C., R. C. Williams, et al. (1988).
Gm35,13,14 and type 2 diabetes mellitus an
association in American Indians with genetic
admixture. Am J Hum Genet 43(4) 520-6.
28
Designs for Family-based LD studies

• Sibling controls (discordant siblings)
• Case-parent trios
• Nuclear families
• Extended families

Ancestral Population
X
Disease
Genetic marker
29
Study Designs used for LD mapping

• Family-based Designs for Association Studies
• Advantages
• Not susceptible to confounding due to population
  substructure
• Tests for linkage and association
• Can test for parent-of-origin effects
• Disadvantages
• Inefficient recruitment, only heterozygous
  parents informative
• Often cannot test for environmental main-effects
• Family members often not available (eg,
  late-onset diseases)

30
TDT (transmission-disequilibrium test)

• Basic idea of TDT
• Disease alleles are transmitted from parents to
  offspring
• Marker alleles in LD with these alleles will also
  be transmitted preferentially to affected
  offspring
• Test if heterozygous parents transmit a
  particular marker allele to affected offspring
  more frequently than expected
• Looks for excess transmission of particular
  alleles from parents to affected children
• Controls are non-transmitted alleles

31
Population stratification

• Random panels of SNPs can be used to test for
  population sub-structure (many new methods).
• First studies - little empirical evidence of
  stratification in large samples from North
  America, Japan, Latin America and Europe.
• Problems with Trios and TDT
• Inefficient in genotyping and sampling
• Difficult (or impossible) to collect
• Can only really have a big problem if doing poor
  epidemiology.
• Potential for bias has been greatly exaggerated.
• Fear of population stratification led to
  substantial changes in study design and analytic
  methods and widespread adoption of trios design
• Now no reason to adopt a family-based design
  solely to protect against stratification

Cardon LR, Palmer LJ. Population stratification
and spurious allelic association. The Lancet
2003361598-604.
32
Why Case/Control?

• Advantages
• Methodology is well-known
• Convenient to collect
• Common
• Very large samples
• More efficient recruitment than family-based
  sampling
• Simultaneous assessment of disease allele
  frequency, penetrance, and AR
• Unrelated controls can provide increased power

• Limitations
• 1. Possible Population Stratification
• 2. Need for highly dense marker sets (capture LD)
  
• Lack of phase information
• Lack of consistency of results

These can be overcome! 1. Assessment and genomic
control of stratification 2. SNP maps 3.
Imputed haplotypes
33
Sample size requirements for case-control
analyses of SNPs (2 controls per case
detectable difference of OR ?1.5 power80).
Statistical power an increasing concern
Palmer, L. J. and W. O. C. M. Cookson (2001).
Using Single Nucleotide Polymorphisms (SNPs) as
a means to understanding the pathophysiology of
asthma. Respiratory Research 2 102-112.
34
Growing utilization of population-based designs

• Increasingly apparent for many diseases that
  population-based studies of unrelated
  individuals, in which case-control and cohort
  studies serve as standard designs for genetic
  association analysis, may be a practical and
  powerful approach
• Power
• Ease and efficiency of collection
• Cohort design longitudinal data and prospective
  assessment
• Birth cohorts
• Value of historical cohorts
• e.g., Nurses Health Study 120,000 Nurses
  recruited in mid-70s 98 follow-up of living
  cohort to current day.
• Pharmacogenetics
• GxE studies

35
An example UK Biobank

• Focus on binary outcomes and gene environment
  interactions
• Large cohort 500,000 individuals
• Age range at recruitment 45-64 yrs ? 45-69 yrs
• Comprehensive exposure assessment at recruitment
• Lifestyle factors, environmental exposures
• Personal and family history of health and disease
• Subsequent monitoring via NHS information systems
• Power set on number of events in 10 years
• ? 40M-60M

Others EPIC, ISIS, Million Women Study
36
Failure to replicate association

• Failure to replicate genetic association studies
  is a problem of genuine concern
• But - more often involves poor study design and
  execution, in particular a lack of appreciation
  for the sample sizes required to detect modest
  genetic effects and over-interpretation of
  marginal results, than undetected population
  stratification.
• Complex human diseases
• Initial detection and replication will likely be
  very difficult.
• multiple testing, laboratory/measurement error,
  and positive publication and investigator-reportin
  g biases
• Population stratification is only one (and
  possibly amongst the least) of many possible
  reasons for non-replication of association
  results.

Cardon LR, Palmer LJ. Population stratification
and spurious allelic association. The Lancet
2003361598-604.
37
Study reproducibility risk allele frequency

• Risk allele frequency has a much greater impact
  on power than disease prevalence for allele
  frequencies lt.2 and gt.8

38
Outstanding Genetics Issues

• Extent of linkage disequilibrium across the
  genome
• Age of expected mutations in populations
• Effect of selection on specific genes
• Allele frequency differences between populations
• Population demographics
• Whether or when subtle population differences are
  important factors in association studies
• Effect of stratification on haplotypes

39
Ultimate Goal