mcmahon-thesis.pdf - stanford university

134
RESEARCH SYNTHESIS FOR MULTIWAY TABLES OF VARYING SHAPES AND SIZE A DISSERTATION SUBMITTED TO THE DEPARTMENT OF STATISTICS AND THE COMMITTEE ON GRADUATE STUDIES OF STANFORD UNIVERSITY IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY onal McMahon November 2009

Upload: khangminh22

Post on 04-Feb-2023

1 views

Category:

Documents


0 download

TRANSCRIPT

RESEARCH SYNTHESIS FOR MULTIWAY TABLES OF

VARYING SHAPES AND SIZE

A DISSERTATION

SUBMITTED TO THE DEPARTMENT OF STATISTICS

AND THE COMMITTEE ON GRADUATE STUDIES

OF STANFORD UNIVERSITY

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

FOR THE DEGREE OF

DOCTOR OF PHILOSOPHY

Donal McMahon

November 2009

c© Copyright by Donal McMahon 2010

All Rights Reserved

ii

I certify that I have read this dissertation and that, in my opinion, it

is fully adequate in scope and quality as a dissertation for the degree

of Doctor of Philosophy.

(Trevor Hastie) Principal Adviser

I certify that I have read this dissertation and that, in my opinion, it

is fully adequate in scope and quality as a dissertation for the degree

of Doctor of Philosophy.

(Robert Tibshirani)

I certify that I have read this dissertation and that, in my opinion, it

is fully adequate in scope and quality as a dissertation for the degree

of Doctor of Philosophy.

(Wing Wong)

Approved for the University Committee on Graduate Studies.

iii

iv

Abstract

This thesis will present techniques for synthesizing partially classified contingency

tables with complex missing data patterns. Data of this form is prevalent in modern

genetics, with disparate research groups performing independent association studies.

We will propose models for combining the results of such studies in a single meta-

analysis.

Two main algorithms are developed in this dissertation. The first is a likelihood-

based approach, using the EM algorithm and loglinear models. Secondly, we will

propose a Bayesian alternative, utilizing the data augmentation algorithm and con-

strained Dirichlet-Multinomial distributions. These general models will then be ex-

tended to deal with data-specific problems; such as retrospective sampling, condi-

tional slices and multiple perspective linked tables. Variance estimation techniques,

model-selection criteria and tests for homogeneity are also derived.

Mendelian diseases are deterministic in nature, with direct genetic inheritance

paths established between parent and offspring. However, the vast majority of in-

herited diseases are in fact non-Mendelian, such as early-onset Alzheimer’s, psoriasis,

breast cancer and cystic fibrosis. Here both genetic and non-genetic factors affect

inheritance patterns, with multiple genes and environmental factors interacting in a

complex fashion. We shall propose methods for the amalgamation of existing clinical

research for such diseases. Each study incrementally measures a particular factor or

group of factors, but is missing data on the combination of all potentially relevant

variables, thereby producing underdetermined results. By integrating these studies

into a single meta-analysis, disease prediction can be carried out across the full set of

risk factors.

v

Acknowledgments

I would like to thank Professor Hastie for his unending support and patience through-

out my PhD. It has been an immensely enjoyable experience to complete this work

under his guidance, especially the early morning surf sessions and statistical chats be-

tween sets. Gene Security Network posed the initial problem and kindly supplied the

datasets in this thesis. Professor Olkin provided much sage advice on meta-analysis

methods and my thesis committee of Professors Tibshirani, Owen, Wong and Lavori

supplied many helpful ideas for the extension of this research. My classmates and the

members of the Hastie-Tibshirani research group also contributed valuable feedback

throughout my time at Stanford.

In addition, I thank the trustees of the Ric Weiland Stanford Graduate Fellowship,

National University of Ireland Travelling Studentship and Fulbright Award for their

generous support of this work.

I have been extremely fortunate to have received guidance and positive direction

from many great teachers and professors, especially in my mathematical training.

I certainly would not have come this far without the support of great educators

such as Donie Houlihan and Prof Philip Boland, and I hope to one day continue

their tradition in moulding future generations of Irish statisticians. Finally and most

importantly, I would like to thank my parents, family and friends who have provided

great encouragement throughout my education, little did they know it would take so

long! Mar a deir an seanfhocal, “ Tig maith mor as moill bheag”.

vi

Contents

Abstract v

Acknowledgments vi

1 Introduction 1

1.1 Outline of the problem . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1.2 In vitro fertilization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Gene Security Network (GSN) . . . . . . . . . . . . . . . . . . . . . . 5

1.4 An introduction to the data . . . . . . . . . . . . . . . . . . . . . . . 6

1.5 Previous research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

1.6 Outline of the dissertation . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Likelihood-based Methods 11

2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

2.2 Loglinear models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12

2.3 Fitting loglinear models . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.4 Meta loglinear models . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.5 Modifications to deal with complex data structures . . . . . . . . . . 18

2.6 The ECM Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.7 Investigating the IPF algorithm . . . . . . . . . . . . . . . . . . . . . 21

2.7.1 Case 1: Full information . . . . . . . . . . . . . . . . . . . . . 21

2.7.2 Case 2: Two margins . . . . . . . . . . . . . . . . . . . . . . . 22

2.7.3 Case 3: Multiple margins and higher dimensional tables . . . . 24

2.8 Testing homogeneity and detecting aberrant studies . . . . . . . . . . 25

vii

2.9 Modifications for retrospective studies . . . . . . . . . . . . . . . . . . 26

2.10 Model selection and testing goodness-of-fit . . . . . . . . . . . . . . . 26

3 Data Augmentation 29

3.1 The Data Augmentation algorithm . . . . . . . . . . . . . . . . . . . 29

3.2 Dirichlet-Multinomial conjugate pair . . . . . . . . . . . . . . . . . . 31

3.2.1 The Multinomial distribution . . . . . . . . . . . . . . . . . . 31

3.2.2 The Dirichlet distribution . . . . . . . . . . . . . . . . . . . . 31

3.2.3 The conjugate pair . . . . . . . . . . . . . . . . . . . . . . . . 32

3.3 Existing models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.3.1 Multinomial saturated model . . . . . . . . . . . . . . . . . . 33

3.3.2 Bayesian constrained model . . . . . . . . . . . . . . . . . . . 34

3.4 Extensions to the DA algorithm . . . . . . . . . . . . . . . . . . . . . 35

3.5 Simulation studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

4 Variance Estimation 39

4.1 The sandwich estimate . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.2 Extending the sandwich estimate to missing data . . . . . . . . . . . 43

4.3 Supplemented EM . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

4.4 The jackknife . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4.5 The bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.6 Bayesian posterior . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

4.7 Simulation studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

5 Retrospective Adjustment 54

5.1 Description of the problem . . . . . . . . . . . . . . . . . . . . . . . . 54

5.2 Maximum likelihood method . . . . . . . . . . . . . . . . . . . . . . . 55

5.3 Mantel-Haenszel method . . . . . . . . . . . . . . . . . . . . . . . . . 55

5.4 Pooling log-odds ratios . . . . . . . . . . . . . . . . . . . . . . . . . . 56

5.5 Modification for retrospective sampling . . . . . . . . . . . . . . . . . 57

5.6 Extension of the modification for retrospective studies . . . . . . . . . 58

5.7 Loglinear-logit model connection . . . . . . . . . . . . . . . . . . . . . 60

viii

5.8 Modification in the loglinear setting . . . . . . . . . . . . . . . . . . . 61

5.9 Simulation studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

6 Psoriasis Meta-Analysis 67

6.1 Psoriasis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

6.2 The data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

6.3 Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

6.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

6.4.1 Model fitting . . . . . . . . . . . . . . . . . . . . . . . . . . . 75

6.4.2 Testing homogeneity and finding influential studies . . . . . . 76

6.4.3 Comparison against standard meta-analysis . . . . . . . . . . 78

6.4.4 Disease prediction . . . . . . . . . . . . . . . . . . . . . . . . . 80

7 Alzheimer’s Disease Meta-Analysis 82

7.1 Alzheimer’s disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

7.2 The data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

7.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87

8 Conclusions 91

A Studies in the psoriasis data base 93

B Interactions present in psoriasis studies 94

C Studies in the Alzheimer’s data base 95

D Interactions present in Alzheimer’s studies 97

ix

List of Tables

1.1 Tsuang et al. [2005] . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1 Poisson Param’s(µ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.2 Observed Data (D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.3 Poisson Param’s(µ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

2.4 Observed Data (D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

5.1 Combining S fully observed two-way tables . . . . . . . . . . . . . . . 54

5.2 Equivalent loglinear and logistic models for a three-way contingency

table with a binary response variable Y . . . . . . . . . . . . . . . . . 61

6.1 Raw Data: Alenius et al. (2002) . . . . . . . . . . . . . . . . . . . . . 71

6.2 Processed Data: Alenius et al. (2002) . . . . . . . . . . . . . . . . . . 72

6.3 G2 and residual deviances for 12 candidate loglinear models . . . . . 75

6.4 Estimated marginal disease probabilities and odds-ratios . . . . . . . 80

7.1 Raw Data: Lehtovirta et al. (1996) . . . . . . . . . . . . . . . . . . . 85

7.2 Processed Data: Lehtovirta et al. (1996) . . . . . . . . . . . . . . . . 86

x

List of Figures

1.1 Four stages in embryonic development, from a single cell to embryo

transfer. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

1.2 GSN Logo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 GSN Process Overview, . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.1 Examples of three-way contingency tables . . . . . . . . . . . . . . . 12

2.2 Further examples of three-way contingency tables, here with slices of

information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

3.1 Simulation results comparing the marginal parameter estimates under

the likelihood-based approach and the Bayesian methods . . . . . . . 38

4.1 Variance estimation: simulation results for cells 1 to 4 . . . . . . . . . 50

4.2 Variance estimation: simulation results for cells 5 to 8 . . . . . . . . . 52

4.3 Investigating alternatives, one sample . . . . . . . . . . . . . . . . . . 53

5.1 Confirming the retrospective adjustment for loglinear models with vary-

ing sample size, as sample size increases both the retrospective and

prospective models provide similarly better estimates . . . . . . . . . 64

5.2 Confirming the retrospective adjustment for loglinear models with vary-

ing the number of studies. . . . . . . . . . . . . . . . . . . . . . . . . 65

5.3 Misspecification of the population disease rate . . . . . . . . . . . . . 66

6.1 Multiple tables from a single study . . . . . . . . . . . . . . . . . . . 73

6.2 Multiple studies to produce the full table . . . . . . . . . . . . . . . . 74

xi

6.3 Convergence for 20 elements of the Psoriasis estimated table . . . . . 77

6.4 Finding studies of high influence using the jackknife influence . . . . . 78

6.5 Combining 5 studies using fixed and random effect models . . . . . . 79

6.6 Prediction intervals for four patients with different risk characteristics 81

7.1 ApoE and NOS3 structures . . . . . . . . . . . . . . . . . . . . . . . 85

7.2 Finding aberrant studies in the Alzheimer’s disease data set . . . . . 87

7.3 Estimated marginal distributions gender and the three genetic risk factors 89

7.4 Prediction intervals for two patients with specific risk characteristic loci 90

xii

Chapter 1

Introduction

1.1 Outline of the problem

At present there is a lack of suitable statistical models that successfully characterize

the effects of genetic and non-genetic variability on non-Mendelian disease risk. For

most of these diseases, there exists no single clinical study which considers all of

the relevant risk factors. Generally there are hundreds of published studies that

investigate a single gene and its association to a particular disease phenotype. Each

study measures a specific factor or a group of factors, but these are merely a subset

of all possibly relevant factors. All is not lost however, as each of these studies

does contain useful information on its respective factors. In this dissertation we will

research new methods to combine multiple clinical studies in a statistically coherent

fashion. This class of problem is known as a meta-analysis, with techniques for

amalgamating, summarizing, and reviewing previous quantitative research in order

to increase statistical power. Care must be taken to adjust for potential publication

bias and to ensure that only homogeneous studies are included in the analysis.

In this research we develop and implement new meta-analysis techniques for deal-

ing with multiway tables arising from multiple studies. The underdetermined data

problem is not novel in statistical research, nor even within meta-analysis. Histori-

cally, techniques such as data imputation, Buck’s method and complete case analysis

have been employed [Cooper and Hedges, 1994]. These standard methods have severe

1

2 CHAPTER 1. INTRODUCTION

limitations, especially in cases where the data is sparse. More advanced model-based

techniques, such as data augmentation and likelihood factorization, require that at

least some complete samples are observed [Fuchs, 1982]. There are many other issues

specific to this type of data set, such as retrospective sampling and multiple slices of

information from a single study, which require novel solutions also. Therefore existing

modeling techniques are not sufficient to solve this problem.

We have developed two distinct approaches to this class of problems. The first of

these is a likelihood-based approach based primarily on the expectation maximization

(EM) algorithm [Dempster et al., 1977] and loglinear models. The EM algorithm

enables maximum likelihood estimation of parameters in probabilistic models, where

the model depends on unobserved latent variables. This allows for model fitting,

even in cases of missing data such as ours. The second method is a Bayesian alterna-

tive, extending data augmentation techniques introduced in Tanner and Wong [1987].

These Bayesian methods allow us to consider the full joint probability distribution.

Using the predictive models developed directly in this research, genetic screening

can be carried out to assess disease probability. Potential applications include the

identification of high-risk patients for preventative care and as a non-intrusive prenatal

testing alternative to amniocentesis. Future clinical research may also be directed by

the evidence garnered from the results of our meta-analysis, further enhancing the

understanding of the disease mechanism. This data may in turn be incorporated in

an advanced second-generation meta-analysis model.

This research grew from a collaboration with a local start-up company. The

company, Gene Security Network (GSN), required the analysis of complex datasets

in order to produce predictive models for a variety of genetic diseases (39 diseases

in total). These models are to be used to enable clinicians and parents make more

informed decisions during the process of in vitro fertilization (IVF). More information

on IVF and preimplantation genetic diagnosis (PGD) is provided in Section 1.2, while

further details on the role of GSN is available in Section 1.3.

1.2. IN VITRO FERTILIZATION 3

1.2 In vitro fertilization

IVF is a fertility treatment in which the female eggs are fertilized outside the woman’s

womb. Following ovarian stimulation, eggs (ova) are removed from the patient’s

ovaries and sperm is added to them in a fluid medium, in vitro. The “best” fertil-

ized egg/eggs (zygotes) are then transferred to the female’s uterus via a thin plastic

catheter, hopefully leading to a completed and safe pregnancy.

While the first successful IVF treatment was achieved in 1978, it is in recent years

that it has exploded in popularity. Today over 1% of births in the United States are

conceived in-vitro, while in Europe rates can be as high as 4% in some countries such

Denmark. It is estimated that infertility affects approximately 6.1 million people in

the USA, with 10% of women of reproductive age having an infertility-related medical

appointment in the past. IVF is the most popular and successful form of assisted re-

productive technology (ART), accounting for 99% of successful births. Births through

IVF have been shown to have over twice the rate of genetic disease and a higher risk

of certain birth defects [Reefhuis et al., 2009]. This high profile study has been highly

cited by those on both sides of the ethical debate regarding genetic screening and

IVF.

Preimplantation genetic diagnosis (PGD) refers to procedures performed on em-

bryos prior to implantation, screening for particular genetic diseases. It currently

offers prospective parents with a family history of a Mendelian genetic disorder, such

as Tay Sachs or Fanconi’s Anemia, the opportunity to avoid passing the disease on to

their children through embryo selection. Utilizing polymerase chain reaction (PCR)

technology, the first screening took place in 1990. It is used in conjunction with IVF

treatment and as an alternative to more invasive prenatal testing techniques such as

amniocentesis. 4-6% of all IVF cycles in the U.S. include PGD, and this number is

growing at 33% per year. In 2005, clinicians performed roughly 134,000 cycles of IVF

in the United States and 653,000 cycles abroad, corresponding to 8,040 and 39,180

cycles of PGD respectively.

• Figure 1.1a shows a single naked oocyte/egg, stripped of the surrounding gran-

ulosa cells.

4 CHAPTER 1. INTRODUCTION

(a) Single cell (b) ICSI (c) Blastomere (d) Blastocyst

Figure 1.1: Four stages in embryonic development, from a single cell to embryo trans-fer.

• In Figure 1.1b an oocyte is injected with sperm during intracytoplasmic sperm

injection (ICSI), a recent technique used in cases of male infertility.

• Figure 1.1c shows the process three days later, at the 8-cell/blastomere stage.

PGD may be performed at this point or at the fourth stage. Previously embryos

at this stage were transferred to the uterus at this point.

• Figure 1.1d shows the more developed blastocyst 2-3 days later in the cycle,

at which point transfer to the uterus is carried out. Blastocyst stage transfers

have been shown to result in higher pregnancy rates.

In oocyte retrieval more than one egg is taken from the ovary, to increase the

chances of forming healthy embryos. However due to risks associated with multiple

births, there are restrictions on the number of embryos transferred to the patient’s

uterus. In the UK, Australia and New Zealand a maximum of two embryos are trans-

ferred, except in unusual circumstances. This decision is based on the individual

fertility diagnosis by the clinician in the US. Currently the choice of embryo is per-

formed by the embryologist based on number of cells, evenness of growth and degree

of fragmentation. PGD can be carried out on embryos who reach the day 3 cell stage

and may be used in combination with other embryonic characteristics to inform the

embryologist’s decision.

1.3. GENE SECURITY NETWORK (GSN) 5

1.3 Gene Security Network (GSN)

GSN Mission: Enable clinicians to use complex genetic and phenotypic information

to make effective medical interventions.

Figure 1.2: GSN Logo

Gene Security Network is a molecular diagnostics company that has developed

proprietary bioinformatics technologies for complex testing of small quantities of ge-

netic material. GSN operates a laboratory for preimplantation genetic diagnosis to

guide doctors in screening embryos for disease susceptibility during in-vitro fertiliza-

tion. The company is based out of Redwood City, California.

Current PGD technology cannot provide parents information on the vast major-

ity of inherited diseases, which have the non-Mendelian inheritance patterns outlined

previously. GSN’s major advancement thus far is in their proprietary technology,

Parental Support, which uses noisy genetic measurements on the blastomeres in com-

bination with:

• Parents’ diploid blood samples

• Father’s haploid sperm sample

• Data from dbSNP

• Data from Hapmap project

Hence they can reconstruct the embryonic DNA at high confidence level; with

sensitivity and specificity above 99.9%. The cost for GSN’s method is $250 versus

the current cost of $5000-$7000 for existing PCR-based techniques. An overview of

the full role played by GSN in a typical IVF treatment is found in Figure 1.3.

If the prospective parents request PGD, blood and sperm samples are taken firstly.

This allows GSN to predict the risk of each genetic disease, utilizing the methods

6 CHAPTER 1. INTRODUCTION

Figure 1.3: GSN Process Overview,

proposed in this dissertation. Based on these results the parents then decide whether

to undertake IVF treatment, and whether to screen the embryos for the diseases for

which they are at risk. It is important to isolate the highest risk diseases, as the

screening of embryos requires removing a single cell at the 8-cell stage, and so it is

not feasible to screen for all diseases. These embryos are then screened using PGD,

aiding the choice for transfer. The statistical models developed here will also be used

at this stage of this process.

1.4 An introduction to the data

In this project there were 39 diseases under consideration. We considered only diseases

where genetic variations have a penetration of more than 50%. A database was

built for each of these diseases, via an extensive literature search and in concordance

with the guidelines for research synthesis outlined in Stroup et al. [2000]. These

protocols include strict rules on the outlining of hypotheses, literature search strategy,

graphical reporting, estimation of publication bias and the provision of guidelines for

1.4. AN INTRODUCTION TO THE DATA 7

future research. Alzheimer’s, cystic fibrosis, breast cancer, myocardial infarction and

psoriasis are some of the candidate diseases under research. In this dissertation we will

concentrate on the analysis of two particular datasets, those relating to Alzheimer’s

disease and psoriasis.

As a general introduction to the data structure however, each database consists

of approximately 100 published papers. Patient recods are aggregated by key demo-

graphic, clinical and genotypic variables. For example in the Alzheimer’s database,

one such paper is Tsuang et al. [2005] (Table 1.1). A summary of the subjects involved

in this study is provided below:

• Gene = ApoE

• Ethnicity = Caucasian

• Familial History = NA

• Onset = NA

• Mean(Age) = 67.70

• SD(Age) = 10.7

There are some variables present (case-control and gender/ApoE), missing (familial

history, onset and ApoE/gender) and conditional (ethnicity) in each of the respective

observed tables.

Case ControlMale 19 93

Female 38 104

Case Controlε2 ε2 0 0ε2 ε3 3 32ε2 ε4 1 3ε3 ε3 21 118ε3 ε4 29 47ε4 ε4 5 3

Table 1.1: Tsuang et al. [2005]

8 CHAPTER 1. INTRODUCTION

1.5 Previous research

Historically, techniques such as (i) the analysis of complete cases only, (ii) single value

imputation and (iii) Buck’s method have been utilized for the meta-analysis of data

with missing values. Under the analysis of only complete cases, it is assumed that

the complete cases are representative of the original sampling. This is not always

reasonable, especially in cases where there is informative censoring/missingness. Sin-

gle value imputation fills in with the mean value of the variable calculated from the

cases that observed the variable. It does however, assume a high degree of homo-

geneity and thus underestimates the variance. Adjustments are possible, but tests

for homogeneity of effect sizes are not. Buck’s method replaces missing values with

the conditional mean. For every pattern of missing data, complete cases are used

to calculate regression equations predicting a value for each missing variable using

the set of completely observed variables. This assumes that the missing variables are

linearly related to other variables in the data.

More advanced model-based techniques have been developed, which may be amended

to deal with categorical data structures for research synthesis. Maximum likelihood

approaches are outlined in Little and Rubin [2002] and analogous Bayesian methods

are provided in Schafer [1997]. Unfortunately, neither of these methods is sufficient

for the needs of the datasets explored in this analysis.

It has been established that there are three different types of missing data in

research synthesis; missing studies in the sample (publication bias), missing effect

sizes from particular studies and missing information on study characteristics. It is

the second of these which is most prevalent in our research. The technical reasons for

missing data in studies are threefold also.

1. Missing completely at random (MCAR): missingness patterns are completely

unrelated to the data itself,

f(M |Y, φ) = f(M |φ) ∀Y, φ (1.1)

where M is the missingness pattern, Y is the complete data set and phi is the

1.6. OUTLINE OF THE DISSERTATION 9

unknown parameter under investigation. If the reasons for the missing values

are not related to any information in the data set itself, then complete cases

may be considered a random sample of the original set of studies.

2. Missing at random (MAR): missingness patterns are related to the completely

observed components,

f(M |Y, φ) = f(M |Yobs, φ) ∀Ymiss, φ (1.2)

where Y = Ymiss ∪ Yobs, with Ymiss and Yobs the missing and observed data

respectively. This assumption is less strict than MCAR and is the most common

made in developing new methods for handling missing data.

3. Not missing at random (NMAR): missingness patterns are related to the missing

values themselves,

f(M |Y, φ) = f(M |Yobs, Ymiss, φ) ∀φ. (1.3)

NMAR would occur if study results or effect sizes were not reported when

not significant. Censoring in survival analysis is another common example of

not missing at random, as patients may leave the study for reasons directly

attributable to the treatment effect.

1.6 Outline of the dissertation

Chapter 2 will consider likelihood-based approaches to this class of problems. We

will introduce the notation used throughout this thesis and derive a generalized EM

algorithm based on meta loglinear models. Extensions and modifications to this

algorithm are also introduced, to deal with issues specific to the data structures.

Tests for homogeneity and finding influential studies are explained, and existing tests

for model adequacy are to accommodate this new class of model.

In Chapter 3, we will introduce a Bayesian alternative based on data augmentation

techniques. In addition to providing a natural method for variance estimation, the

10 CHAPTER 1. INTRODUCTION

derived algorithm will allow for the analysis of the full posterior distribution.

Chapter 4 concentrates on the various options available for variance estimation

under the models proposed in earlier chapters. We will consider such methods as

the sandwich estimator, the jackknife, bootstrapping, posterior standard error and

multiple imputation. Simulation studies are carried out to establish the adequacy of

each of these methods.

In Chapter 5, we will develop methods to adjust for retrospective sampling in mul-

tiway tables. Logistic and loglinear models are compared, with instances of equiva-

lency and difference investigated. Simulations studies confirm the validity of the new

retrospective adjustments.

Chapter 6 and Chapter 7 will contain the results of the analysis of the Alzheimer’s

and psoriasis data sets. We will fit both the likelihood-based and Bayesian models

and investigate the model adequacy, comparing against the limited existing methods.

Discussion and conclusions shall be provided in Chapter 8.

Chapter 2

Likelihood-based Methods

2.1 Notation

So if we consider a multivariate distribution π obtained by the crossing of a collection

of K categorical factors F = F1, . . . FK, the kth of which has Lk levels. π is a

multiway table of probabilities with each element ∈ [0, 1] and the sum of all the

elements is 1. The dimension of π is L1 × L2 × . . . × LK . We will use the notation∑F πF = 1. If we partition the variables in F into two mutually exclusive subsets

O and M, with O ∪M = F , then πO =∑M πF =

∑M πO,M denotes the marginal

table indexed by variables in O obtained by summing the entries of π over all levels

of the variables inM. O shall be referred to as observed variables andM as missing

or marginal variables.

We have data from S different studies, and the ith such study gives us an observed

table NOi, i.e it is a complete table on a subset Oi of the variables in F . The Oi of

different studies will typically involve different variables, and also different numbers

of variables. Also, typically none of the studies will have Oi = F , although this is not

excluded. The goal of this meta-analysis is to combine all these studies to produce a

coherent estimate πF of πF .

We shall also generalize this model to deal with other kinds of partial information:

(i) Rather than a marginal table we sometimes see a section or slice; we see a

complete table in Oi, but rather than marginalized wrt toMi, it is conditioned

11

12 CHAPTER 2. LIKELIHOOD-BASED METHODS

(a) Basic three factor contingency table (b) Two factors observed, one marginal/missing

Figure 2.1: Examples of three-way contingency tables

on particular values for each of the variables in Ci, with Oi ∪ Ci = F .

(ii) We can see both marginals and slices. We see a complete table inOi, conditioned

on particular values of variables in Ci, and all marginalized wrt to Mi, with

Oi ∪ Ci ∪Mi = F . A figurative example of this is shown in Figure 2.2a.

(iii) It is possible for a single study to comprise of multiple tables, each with their

own set of observed, missing and conditional variables. Study i may contain of

numerous tables, the jth of which has the following variable set (Oji ,Mji , C

ji ),

with Oji ∪Mji ∪C

ji = F for each j. Multiple colored slices may be seen in Figure

2.2b.

Initially we deal only with the simple marginal case, but later we discuss these other

three cases also.

2.2 Loglinear models

A traditional approach to modeling a multiway table is to represent the probabilities

by a loglinear model log π = η, where we implicitly assume that the entries in π are

strictly ∈ (0, 1). Usually we have only a single observed table N , and impose structure

2.2. LOGLINEAR MODELS 13

(a) Example of a slice (b) Multiple slice example

Figure 2.2: Further examples of three-way contingency tables, here with slices ofinformation

on the table by restricting η to have an ANOVA representation wrt the factors. So

for example, if F = F1, F2, F3, then the loglinear model

log πF = ηF1 + ηF2 + ηF3 (2.1)

represents a model in which the probabilities for the three-way table are products of

three terms, one corresponding to each factor. This corresponds to the full indepen-

dence model for the three dimensional distribution represented by π. Likewise,

log πF = ηF1,F2 + ηF3 (2.2)

represents a model with independence between F1, F2 and F3, but dependence

between F1 and F2.

This notation is still abstract; in reality for this example we will need to represent

specific entries in the table, such as πijk. This is the probability of seeing (F1 =

i, F2 = j, F3 = k). In this case the notation in (2.1) implies

log πijk = ηiF1+ ηjF2

+ ηkF3. (2.3)

14 CHAPTER 2. LIKELIHOOD-BASED METHODS

Thus the number of different constants of the form η`F represented by a generic term

like ηF is the number of levels of F . Likewise, the number of constants for a generic

term ηF1,F2 is L1 × L2.

Just as in multiway ANOVA, this would lead to a redundant coding, and certain

parameters would be aliased with each other and hence not be identifiable. Two

general approaches to combat this are

1. Set every instance of ηLj

Fj= 0 — i.e. any constant involving any of the factors

at the highest level to zero.

2. Include a quadratic regularization term on all the constants when fitting the

model.

For this application we prefer 1.

One can enumerate the entire set of models of this form for any given high-

dimensional table. Typically we chose one that has simple structure, but represents

the observed data well.

As an aside, many of the models correspond to some type of independence or con-

ditional independence, and hence can be represented by a graphical model (directed

acyclic graph). There are some, such as

log πF = ηF1,F2 + ηF2,F3 + ηF1,F3 (2.4)

(no third-order interaction model) which does not represent any form of conditional

independence, and cannot be uniquely represented by a graphical model.

Usually we represent model such as (2.1),(2.2) & (2.4) in terms of a model matrix

X and a parameter vector θ:

log π = η(θ) = Xθ (2.5)

Here π is a vector of probabilities of length∏K

k=1 Lk, filled in lexicographical ordering

(indices varying most rapidly from right to left). The parameter vector θ consists of

all the identifiable parameters in the model, excluding the ones that are zero, and

2.3. FITTING LOGLINEAR MODELS 15

the rows of X are filled with zeros and ones to indicate the presence or absence of a

particular parameter for that element of log π.

Loglinear models are well described in a number of books, such as McCullagh and

Nelder [1983].

2.3 Fitting loglinear models

Typically loglinear models are fit using Poisson maximum-likelihood. Often a multi-

nomial is more appropriate, since the original sample was conditional on certain

marginals. It turns out that as long as there are terms in the loglinear model corre-

sponding to these fixed counts, Poisson ML is equivalent to multinomial ML.

The log-likelihood of an observed table, given a model structure is

`(θ) = n∑F

(rFηF(θ)− eηF (θ)), (2.6)

where rF = NF/n are the observed proportions, and n =∑F NF is the total count in

the table. This log-likelihood is convex in θ (if X is full column rank). Differentiating

wrt θ, and using (2.5), we get (in matrix notation)

d`(θ)

dθ= nXT (r− π) = 0 (2.7)

These equations are quite intuitive, since X is binary. It says that certain marginals

of the fitted table π should match the corresponding data marginals. In fact, the

marginals that have to match correspond exactly to the presence of terms indexed

by factors in (2.1),(2.2) & (2.4). The iterative proportional fitting algorithm (IPF)

exploits this fact, and starting with a constant table, cycles around correcting the

table so that it matches each marginal as required in turn.

Alternatively we can compute the Hessian matrix

d2`(θ)

dθdθT= −nXTDπX, (2.8)

16 CHAPTER 2. LIKELIHOOD-BASED METHODS

and use the Newton algorithm to solve for θ. Here Dπ = diag(π).

Conveniently, the Newton algorithm can be represented as an iteratively reweighted

least squares (IRLS) algorithm:

1. Compute the working response z = η + D−1π (r− π).

2. Fit a weighted linear regression of z on X with weights Dπ to update the

coefficients θ.

2.4 Meta loglinear models

We now propose a method to generalize the loglinear model for the multiple study

scenario outlined in Section 2.1. Each of the observed tables N iOi

is indexed by a

subset Oi ⊆ F of the full collection of factors. We consider the following model for

π:

log πF =S∑i=1

ηOi. (2.9)

This model has loglinear terms to cover each of the observed tables, likely with many

redundancies. These redundancies can easily be removed when the model is repre-

sented in the form (2.5), simply by removing duplicate columns in X. We will write

this model as

log πF = xTFθ (2.10)

We propose to fit the model by maximizing the likelihood of the observed tables

N iOi

. Oi represents the factors in F observed for study i, and its complementMi are

those factors in F not observed. The probabilities under the model of the observed

factors are

πOi=

∑Mi

πF (2.11)

=∑Mi

exTFθ (2.12)

2.4. META LOGLINEAR MODELS 17

Hence the sum of the Poisson log-likelihoods of the observed tables is

`(θ) =S∑i=1

ni∑Oi

[riOilog πOi

(θ)− πOi(θ)] (2.13)

Again riOiare the observed proportions corresponding to N i

Oi, and ni =

∑OiN iOi

is

the total count for study i. As such ni is the weight assigned to study i, and we may

consider other weights if there is too much imbalance.

Although in principle we could go through the motions to maximize (2.13), we

no longer get a simple expression for the gradient. This is because each of the terms

log πOi(θ) is a log of a sum of exponential terms, and does not simplify. This is

a classical case for the EM algorithm [Dempster et al., 1977], which is an iterative

algorithm for simplifying such situations.

Next we present the EM algorithm for this meta analysis. It consists of alternating

the following two steps till convergence.

E Step: For each observed table riOi, fill it out to become a full table riF by expanding

the missing dimensions using the current estimate πF :

riF = riOiπMi|Oi

(2.14)

= rOi

πFπOi

(2.15)

M Step: Fit the model using the filled out tables by maximizing the full log-likelihood

`full(θ) =M∑i=1

ni∑F

[riF log πF(θ)− πF(θ)]. (2.16)

Note that, because of (2.10), the first term in the sum simplifies. It is easy to see

that the gradient is given by

d`full(θ)

dθ=

M∑i=1

niXT (riF − πF). (2.17)

18 CHAPTER 2. LIKELIHOOD-BASED METHODS

Letting

rF =

∑Si=1 nir

iF∑S

i=1 ni, (2.18)

we see that the likelihood equation simplifies to

XT (rF − πF) = 0. (2.19)

We are back in the situation of Section 2.3, and this equation can easily be solved by

either the Newton method or IPF.

There may be occasion to fit a saturated multinomial model rather than the meta-

loglinear model outlined in the algorithm above. This may be achieved quite easily

with an EM algorithm similar to that outlined above. The E-step in fact remains

completely unchanged, with the M-step becoming merely a weighted mean of the

expanded tables.

2.5 Modifications to deal with complex data struc-

tures

In the introduction to this chapter we described three types of data structure not

addressed in the basic EM algorithm outlined in Section 2.1. In this section we

propose some amendments to the algorithm to incorporate such data.

(i) We may observe all the variables in Oi, but at fixed levels of each of the variables

in Ci. In this case, we need to modify our model and the E-step of the EM

algorithm. For the model, we should include a term corresponding to Oi ∪Ci =

F ; in other words the complete model. For the E-step, let ci be the actual

levels of the variables in Ci that are observed; hence our observed partial table

can be written ni · riOi|Ci=ci . Let the current estimated conditional table be

πOi|Ci . Let πiOi|Ci be the modification of πOi|Ci obtained by replacing πOi|Ci=ci

with riOi|Ci=ci .Then

riF = πiOi|Ci πCi . (2.20)

2.5. MODIFICATIONS TO DEAL WITH COMPLEX DATA STRUCTURES 19

(ii) If we observe a slice in some variables, and some are missing (marginal), then

our strategy is similar. The model term corresponds to Oi ∪Ci. For the E-step,

we need to first marginalize πF with respect toMi to compute πOi∪Ci and hence

πOi|Ci . Then we proceed as above, obtaining

riOi∪Ci = πiOi|Ci πCi , (2.21)

and finally

riF = riOi∪Ci πMi|(Oi∪Ci) (2.22)

= πiOi|Ci πCi πMi|(Oi∪Ci).

(iii) We may observe multiple slices, each with an associated set of missing variables.

Again, adjustments are required to the E-step and the model term. Firstly we

need to marginalize with respect to the appropriate missing dimensions Mji

for all j = 1, . . . , J , to compute πOji∪C

ji

and hence πOji |C

ji. Proceeding with the

modification steps already outlined above, we can calculate the estimated full

table for the jth table in study i

rij

F = πij

Oji |C

ji

πCjiπMj

i |(Oji∪C

ji ). (2.23)

Hence we can find the estimated full table for study i as the weighted mean of

these tables

riF =J∑j=i

njiniri

j

F . (2.24)

Although not obvious at first, this is in fact equivalent to providing a relative

weighting on the observed partial tables and then carrying out the expansion

in a more step-by-step process. This is explained in Section 6.3. Both methods

produce the ML solution for the full set of expanded tables, as each observed

table contains an independent set of observations. This elegant solution is only

possible since we assumed disjoint perspectives only, which is true in our dataset,

20 CHAPTER 2. LIKELIHOOD-BASED METHODS

but perhaps not more widely. Therefore we mutually satisfy each observed

margin, without introducing any further model complexity. Similarly the P-Step

in the Bayesian method outlined in Sections 3.3.1 and 3.4 provides an equivalent

solution involving the summation of the cell counts rather than weighting the

cell probabilities.

In more complicated situations where we do not have disjoint slices, it is neces-

sary to use the IPF algorithm in solving for the ML estimate of the full table.

This would allow us to mutually satisfy the marginal densities, even if they

contain some intersection. The Bayesian solution would follow a similar line,

with constrained sampling from product multinomial distributions.

There are in fact multiple model terms relating to this study;O1i ∪ C1

i , . . . ,OJi ∪ CJi

.

In each of the three cases above, the weight ni is the total number of observations

observed in the study.

2.6 The ECM Algorithm

In many cases, and especially with large data sets, the EM algorithm may be unduly

cumbersome. Even with the huge advances in computing speed convergence times

may be debilitating, and therefore speed-ups to the algorithm are attractive. It has

been shown [Meng and Rubin, 1993] that it may not be necessary to iterate until full

convergence at each M-step, with a single cycle of the model-fitting process sufficient.

In the context of the algorithm we have proposed in this chapter, this would equate

to a single cycle of the IPF algorithm. This modification has become known as

the expectation-conditional maximization (ECM) algorithm. ECM retains the same

reliable convergence properties as EM, increasing the observed-data log-likelihood at

each step. The M-step is replaced by the quicker CM-step, which still asymptotically

converges to a maximum over the full parameter space ΘM . As in IPF, the starting

values for θ should lie in the interior of the parameter space, with structural zeros

being assigned a zero/null value and uniform values elsewhere as advised in Agresti

[2002].

2.7. INVESTIGATING THE IPF ALGORITHM 21

µ11 µ12 . . . µ1J

µ21 µ22 . . . µ2J

......

. . ....

µI1 µI2 . . . µIJ

Table 2.1: Poisson Param’s(µ)

n11 n12 . . . n1J

n21 n22 . . . n2J

......

. . ....

nI1 nI2 . . . nIJ

Table 2.2: Observed Data (D)

2.7 Investigating the IPF algorithm

Throughout this dissertation we speak about methods to combine different sources

of marginal and conditional information, in fact this is the central thesis of our work.

In particular we have discussed the use of the iterative proportional fitting algorithm

(IPF) as a method to find the ML estimates when we have multiple perspectives on

the same data. This method is well-established in cases where a full table is known

and a constrained model is required [Agresti, 2002]. However, the literature does not

explicitly consider data where multiple independent margins or slices are produced

from a single study, i.e. where no full table is observed. In this section we will

introduce some of the different forms of marginal information that may arise from

Poisson generated contingency tables, where data arrives only in a partially classified

form. We will show that the estimates produced by the IPF algorithm are in fact the

ML estimates of the unknown parameters (cell probabilities).

2.7.1 Case 1: Full information

Firstly and most trivially, we will consider the simplest case whereby we have full

information on the table of interest, in order to familiarize ourselves with the notation

that will be used hereafter. In the two-way example we are given an IxJ table of

Poisson parameters (µ) and an IxJ set of observed counts (D).

Pµ(D) =I∏i=1

J∏j=1

e−µijµnij

ij

nij!(2.25)

= eµ..

I,J∏i,j=1

µnij

ij

nij!(2.26)

22 CHAPTER 2. LIKELIHOOD-BASED METHODS

Therefore the log-likelihood is,

⇒ `D(µ) = −µ.. +

I,J∑i,j=1

nijlogµij +

I,J∑i,j=1

log(nij!) (2.27)

and unsurprisingly the MLE’s are,

µij = nij (2.28)

2.7.2 Case 2: Two margins

In this second case we shall consider situations where we are provided with two

margins of information for a two-way table. We are not privy however to any full

information, i.e. a fully categorized two-way table. We observe the two margins

D = D1 ∩D2, shown in Table 2.4.

µ11 µ12 . . . µ1J µ1.

µ21 µ22 . . . µ2J µ2.

......

. . ....

...µI1 µI2 . . . µIJ µI.

µ.1 µ.2 . . . µ.J µ..

Table 2.3: Poisson Param’s(µ)

n1.

n2.

...nI.

n.1 n.2 . . . n.J

Table 2.4: Observed Data (D)

Pµ(D) = P (D1 ∩D2) = P (D1 ∩D2|n..)P (n..) (2.29)

= P (D1|n..)P (D2|n..)P (n..) (2.30)

2.7. INVESTIGATING THE IPF ALGORITHM 23

P (D1|n..) = P (X1. = n1., . . . , XI. = nI.|x.. = n..) (2.31)

=

∏Ii=1

e−µi.µni.i.

ni.!e−µ..µn..

..

n..!

(2.32)

=n..!∏Ii=1 ni.!

I∏i=1

µni.i.

µn....

(2.33)

=n..!∏Ii=1 ni.!

I∏i=1

πni.i. (2.34)

⇒ Pµ(D) =

(n..!∏Ii=1 ni.!

I∏i=1

πni.i.

)(n..!∏Jj=1 n.j!

J∏j=1

πn.j

.j

)(e−µ..

µn....

n..!

)(2.35)

`D(π, µ..) =I∑i=1

ni.logπi. +J∑j=1

n.jlogπ.j − µ.. + n..logµ.. + . . . (2.36)

∂lD(π, µ..)

∂µ..= −1 +

n..µ..

(2.37)

⇒ µ.. = n.. (2.38)

and the profile likelihood is,

`D(π) =I∑i=1

ni.logπi. +J∑j=1

n.jlogπ.j + . . . (2.39)

This has no direct solution, but if the two contraints,∑I

i=1 πi. = 1 and∑J

j=1 π.j = 1,

are added via Lagrangian multipliers we get a modified profile likelihood,

`D(π, λ1, λ2) =I∑i=1

ni.logπi. +J∑j=1

n.jlogπ.j − λ1

(I∑i=1

πi. − 1

)− λ2

(J∑j=1

π.j − 1

)(2.40)

24 CHAPTER 2. LIKELIHOOD-BASED METHODS

∂`D(π, λ1, λ2)

∂λ1

=I∑i=1

πi. − 1 Constraint 1 (2.41)

∂`D(π, λ1, λ2)

∂λ2

=J∑j=1

π.j − 1 Constraint 2 (2.42)

∂`D(π, λ1, λ2)

∂πi.=

ni.πi.− λ1 (2.43)

⇒ πi. = λ1ni. =ni.n..

(2.44)

with the last equality holding since,

I∑i=1

πi. =I∑i=1

λ1ni. = 1 (2.45)

⇒ λ1 =1∑Ii=1 ni.

=1

n..(2.46)

Since πi. =µi.µ..

it is found that µi. = ni. for all i = 1, . . . , I and similarly µ.j = n.j

for all j = 1, . . . , J . Therefore, given two margins of information, the likelihood only

provides information regarding the marginal sums. The estimates are consistent with

those produced under IPF.

2.7.3 Case 3: Multiple margins and higher dimensional ta-

bles

Generalizing the previous case, we can easily extend the proof to multidimensional

tables (IxJxKx. . .). For example in the IxJxK case we find that the maximum

likelihood approach leads to µi.. = ni.. for all i = 1, . . . , I, µ.j. = n.j. for all j = 1, . . . , J

and µ..k = n..k for all k = 1, . . . , J . Again it is relatively straightforward, using the

methods from Case 2 above, to show equivalency between the estimates provided by

the IPF algorithm and the true ML estimates.

2.8. TESTING HOMOGENEITY AND DETECTING ABERRANT STUDIES 25

2.8 Testing homogeneity and detecting aberrant

studies

Before pooling the estimates of an effect size from a series of studies, it is important

to determine whether the studies can be described as sharing a common effect size.

The Q-statistic [Hedges and Olkin, 1985] has been developed as a statistical test for

the homogeneity of the effect size. Formally it is a test of the hypothesis

Ho : θ(1) = θ(2) = . . . = θ(S) (2.47)

versus the alternative that at least one differs. θ(i) is a length p vector of effect sizes for

study i, with θ(i) the associated estimate. Define the column vector θ∗ of dimension

Sp by

θ∗ = (θ(1), . . . , θ(S))′. (2.48)

Similarly define the associated estimated covariance matrix Σ∗ by

Σ∗ = Diag(Σ(1), . . . , Σ(S)). (2.49)

where Σ(1), . . . , Σ(S) are the large sample estimates of the covariance matrices of the

θ(1), . . . , θ(S). We can now calculate the Q-statistic

Q∗ = θ′∗Cθ∗, (2.50)

where

C = Λ∗ − Λ∗ee′Λ∗/e

′Λ∗e, (2.51)

Λ∗ is the inverse of Σ∗ and e is column vector of Sp ones.

The formal test is based in the fact that if θ(1) = . . . = θ(S) and the sample size

in all studies is reasonably large, then Q∗ has a chi-square distribution with (S − 1)p

degrees of freedom.

Another method to detect aberrant studies utilizes the jackknife samples found

in the variance estimation (Section 4.4). Faulty or suspicious data can be identified

26 CHAPTER 2. LIKELIHOOD-BASED METHODS

using the jackknife influence statistic, which measures the distance (d) between the

leave-one-out estimate and the left-out observations

d(θ(i), θ(.)

). (2.52)

It is necessary to select an appropriate distance measure for multivariate analysis, and

we have chosen to use the established Kullback-Leibler divergence (relative entropy)

[Kullback and Leibler, 1951],

d(θ(i), θ(.)

)=∑

θ(i)logθ(i)

θ(.)

. (2.53)

2.9 Modifications for retrospective studies

The methods proposed thus far in this dissertation have dealt exclusively with data

collected from prospective research studies. However, in retrospective or observational

sampling designs, such as case-control biomedical studies, an adjustment is required

in order to correctly estimate the effect sizes. This topic will be explored detail in

Chapter 5.

2.10 Model selection and testing goodness-of-fit

It is customary to measure the quality of a model and test it against alternatives, to

ensure both model optimization and parsimony. The goodness-of-fit of a statistical

model describes how well it fits a set of observations. For loglinear models we use the

deviance likelihood ratio test (G2) or Pearson chi-squared statistic (X2)

G2 = 2∑j

njlog

(njµj

)X2 =

∑j

(nj − µj)2

µj(2.54)

2.10. MODEL SELECTION AND TESTING GOODNESS-OF-FIT 27

with nj denoting the observed data and µj the expected counts under the proposed

model.

These tests may be extended to cases with partially classified tables. In these

circumstances, we sum over the incomplete tables, but unlike the complete-data cases

we obtain nonzero values for the test statistics under the saturated model. In fact

the values for G2 and X2 for the saturated model provide tests for whether the data

are missing completely at random (MCAR) or missing at random (MAR). Chi-square

statistics for restricted models may be obtained by calculating G2 (or X2) for both

the restricted model and the saturated model, and subtracting these two quantities

[Fuchs, 1982].

G2 = 2S∑i=1

∑Oi

niriOilog

(riOi

rOi

)−G2

0

X2 =S∑i=1

∑Oi

ni(riOi− rOi

)2

rOi

−X20 (2.55)

where G20 and X2

0 denote the value of the statistic evaluated at the MLE for the

saturated model. Also both G2 and X2 are χ2 distributed with df = q− p− 1, where

q is the total number of cells in the contingency table and p is the number of terms

in the fitted loglinear model. It should be noted that these two test statistics have

the same number of degrees of freedom as the chi-square test for the restricted model

with complete data. Using these tests we may compare competing models and test

hypotheses such as the inclusion and exclusion of parameters.

An alternative method for choosing the most appropriate model is cross-validation.

Cross-validation has been proposed in model-selection for many other situation such

as those outlined in Hastie et al. [2001]. We naturally have a leave-one-out sample

via the jackknife method (Section 4.4). For each sample, we fit models of different

sizes to each of the training set, with α denoting the tuning parameter of model size,

28 CHAPTER 2. LIKELIHOOD-BASED METHODS

and then test each model on the left out sample.

CV (α) =1

N

N∑i=1

L(yi, f

−κ(i)(xi, α))

(2.56)

Here CV (α) is an unbiased estimate of the test error curve, under some chosen loss

function L. yi are the observed responses, and f−κ(i)(xi, α) is the estimated fit on the

test set, based on the model found on the training set. Hence we find the model size

α which minimizes this test error. The best model (f(x, α)) of this size is then fitted

to the full data set. It should be noted that the fitted model of size α may or not be

equivalent to any of the best models in the leave-one-out samples.

Chapter 3

Data Augmentation

3.1 The Data Augmentation algorithm

To date we have exclusively considered and developed likelihood-based approaches

to this class of problems. However, various Bayesian methods have been created as

an alternative to the EM framework, most notably data augmentation [Tanner and

Wong, 1987]. Throughout this chapter I will assume that the reader a rudimentary

knowledge of modern statistical methods, and so will not delve into great depth on

the basics of distribution theory nor Bayesian methods.

The data augmentation (DA) algorithm is analogous to the EM algorithm, in

that is exploits the simplicity of the likelihood function (posterior distribution) of

the unknown parameter given the augmented data. Interestingly the steps of the

algorithm also follow the same logic as the EM and this is seen below. In contrast to

the EM algorithm where just the maximum and curvature are found, in DA the entire

posterior distribution is obtained. This is especially useful in improving inference in

small sample cases, where assumptions about the regularity of the likelihood may be

questionable.

The DA algorithm augments the observed data Y with some latent data Z. The

overall aim of this algorithm is the calculation of the posterior distribution p(θ|Y ), but

unfortunately this is intractable due to the presence of the latent data. Given both

Y and Z, it is assumed that one can calculate or at least sample from p(θ|Y, Z), the

29

30 CHAPTER 3. DATA AUGMENTATION

augmented posterior distribution. So in order to procure the posterior distribution,

multiple imputations of Z from the predictive distribution p(Z|Y ) are found and then

we compute the average of p(θ|Y, Z) over these imputations. However since p(Z|Y )

depends on p(θ|Y ), an iterative algorithm is necessary for the calculation of p(θ|Y ).

There are two identities which provide the foundation for the DA algorithm:

1. The posterior identity:

p(θ|Y ) =

∫Z

p(θ|Y, Z)p(Z|Y )dZ. (3.1)

2. The predictive identity:

p(Z|Y ) =

∫Θ

p(Z|φ, Y )p(φ|Y ), (3.2)

where p(Z|φ, Y ) is the conditional predictive distribution. Monte Carlo methods

are used to perform the integration in the posterior identity. Given a value θ(t) of θ

drawn at iteration t the DA algorithm iterates between the following two steps:

Imputation (I) Step: Generate a sample z1, z2, . . . , zm (Z(t+1)) from the current

approximation to the predictive distribution p(Z|Y, θ(t)).

Posterior (P) Step: Update the current approximation of p(θ|Y ) as the mixture of

the augmented posteriors of θ, i.e. draw θ(t+1) with density p(θ|Y, Z(t+1)).

This iterative procedure can be shown to eventually converge to a draw from the

joint distribution of Z, θ|Y as t tends to infinity. The value of m need not be very

large, in fact with m = 1 the DA algorithm reduces to a special case of the Gibbs

sampler where the random vector is just partitioned into two sub-vectors [German

and German, 1984].

3.2. DIRICHLET-MULTINOMIAL CONJUGATE PAIR 31

3.2 Dirichlet-Multinomial conjugate pair

3.2.1 The Multinomial distribution

Suppose that Y = (y1, y2, . . . , yn)T , with yi a categorical variable taking one of C

possible values c = 1, 2, . . . , C. If we set nc to be the number of observations for

which yi = c, then∑C

c=1 nc = n. Conditional on the total sample size n, the counts

in each category (n1, n2, . . . , nC) have a multinomial distribution with probabilities

π = (π1π2, . . . , πC) and index n. It should be noted that∑C

c=1 πc = 1 and therefore

the sampling distribution is:

p(Y |π) =

(n!

n1!n2! . . . nC !

)( C∏c=1

πncc

)(3.3)

Hence we find the likelihood of θ to be:

`(π|Y ) =C∑c=1

nclogπc, (3.4)

and the MLE is found to be πc = nc/n, the sample proportion. The binomial is a

special case of the multinomial distribution, where C = 2.

3.2.2 The Dirichlet distribution

Suppose that π = (π1, π2, . . . , πC) is a vector of random variables with the property

that πc ≥ 0 for all c = 1, 2, . . . , C and∑C

c=1 πc = 1. Then π is said to have a Dirichlet

distribution with parameter α = (α1, α2, . . . , αC) with density:

p(π|α) =Γ(∑C

c=1 αc)

Γ(α1)Γ(α2) . . . ,Γ(αC)πα1−1

1 πα2−12 . . . παC−1

C (3.5)

over the simplex Π, where Γ denotes the gamma function. This is a valid proba-

bility density if πc > 0 for all c = 1, 2, . . . , C.

The Dirichlet distribution is a multivariate generalization of the Beta distribution.

32 CHAPTER 3. DATA AUGMENTATION

3.2.3 The conjugate pair

In fact the Dirichlet density 3.4 is of the same functional form as equation 3.3 and

so they form a conjugate pair. So if we assume that the prior density for the π

parameters in equation 3.3 have a Dirichlet distribution, D(α), then the posterior

distribution is found to be:

p(π1, π2, . . . , πC |Y ) ∝C∏c=1

πnc+αc−1C , (3.6)

again with πc > 0 for all c = 1, 2, . . . , C and∑C

c=1 πc = 1. In other words, it is

Dirichlet(nc + αc). Therefore, the posterior mean of πc is (nc + αc)/(n. + α.), where

n. =∑C

c=1 nc and α. =∑C

c=1 αc. There are some common choices for αc:

1. αc = 0 for all c = 1, 2, . . . , C, here the posterior mean coincides with the ML

estimate for complete-data cases for parameters which are linear functions of the

estimated terms π1, π2, . . . , πC . This choice is not suitable if there are empty

cells in the contingency table. This is an improper prior; the existence of a

proper posterior under this prior is not guaranteed.

2. αc = 1/2 for all c = 1, 2, . . . , C, this yields Jeffreys prior, an improper prior here

but a reasonable compromise between the choices of αc = 0 or αc = 1.

3. αc = 1 for all c = 1, 2, . . . , C, this is a diffuse prior and yields the uniform

distribution.

4. αc > 1 for all c = 1, 2, . . . , C, can be used as a flattening prior for sparse tables.

3.3 Existing models

Data augmentation methods have been developed to deal with structures similar to

those in our datasets. In fact in the original paper Tanner and Wong [1987] there

is some work on latent class analysis which utilizes the Dirichlet-Binomial conjugate

pair. Schafer [1997] elaborates on this area further, introducing two models for data

3.3. EXISTING MODELS 33

similar 1 to ours, the multinomial saturated model 3.3.1 and the constrained Bayesian

model 3.3.2. While the models proposed by Schafer deal with similar but more basic

data structures to those seen in our datasets, much work was required to develop new

methods to deal with such issues as:

• Each study provides information on just a subset of the risk factors. Schafer’s

methods deal with fully classified tables with additional partially classified ta-

bles.

• Dealing with conditional slices, where data is observed at conditional values of

some of the risk factors.

• Combining multiple sources of information from within a single study.

• Adjusting for retrospective sampling in each study.

3.3.1 Multinomial saturated model

Throughout this and Section 3.3.2, it is assumed that the ith observed study table,

riOi, contains information on a subset Oi of the K categorical factors. The remaining

factors, Mi, are missing for that study. In the EM algorithm in Chapter 2, the E

step consisted of filling out each riOiover the missing variables using the appropriate

conditional table, πMi|Oi, from the current estimate of the full table. Analogously in

the DA algorithm we simulate the sampling distribution to produce an appropriate

estimate of the full table for each study, riF . Under the assumption of a Dirichlet

prior θ ∼ D(α), the P step is then just a random simulation of θ from the augmented

posterior D(α + r). The algorithm iterates between the following two steps:

I Step: Draw each riF from its respective product multinomial distribution

M(niriOi, πMi|Oi

). (3.7)

1All data augmentation methods in this thesis were in fact developed independently of Schafer’swork, but the author does wish to acknowledge similarities in the basic methods.

34 CHAPTER 3. DATA AUGMENTATION

where πMi|Oi= πF

πOi. The summation of these complete-data tables r =

∑Si=1 nir

iF

is found for use in the P step below, as each of these simulated tables is viewed

as an independent draw from the true multinomial distribution. Hence, under the

Dirichlet-multinomial conjugacy in Section 3.2.3, the multinomial parameter for each

cell is the sum of the respective cells from the constituent tables.

P Step: Draw πF with from the augmented posterior density

D(r + α) (3.8)

3.3.2 Bayesian constrained model

The Bayesian iterative proportional fitting (IPF) DA algorithm follows much the

same form as that of the saturated multinomial model. In fact the I-steps in both

are exactly equivalent, with the changes coming in the posterior (P) step. Instead

of fitting a saturated multinomial model at the P step, constraints are put on the

Dirichlet posterior, mimicking those of the loglinear models in Section 2.4. The

iterative method of generating random draws from a constrained Dirichlet posterior

was first presented in Gelman et al. [1995]. There are obvious similarities between

this method and iterative proportional fitting; hence it was termed Bayesian IPF.

An example of the algorithm in operation is provided below for a three-way con-

tingency table, fitted with only two-way interactions (the model of homogeneous as-

sociation). The previous P step is replaced by three conditional posterior (CP) steps.

In the algorithm below, each of the r terms is a proportion in the observed tables,

and gijk are the simulated proportions in accordance with the model restrictions as

outlined for example in Equation 3.12.

I Step: Draw each riF from its respective product multinomial distribution

M(niriOi, πMi|Oi

). (3.9)

3.4. EXTENSIONS TO THE DA ALGORITHM 35

CP1 Step:

π(t+1/3)jkl = π

(t+0/3)jkl

(gjk+/g+++

π(t+0/3)jk+

)∀j, k, l. (3.10)

CP2 Step:

π(t+2/3)jkl = π

(t+1/3)jkl

(gj+l/g+++

π(t+1/3)j+l

)∀j, k, l. (3.11)

CP3 Step:

π(t+3/3)jkl = π

(t+2/3)jkl

(g+kl/g+++

π(t+2/3)+kl

)∀j, k, l. (3.12)

Here g(t+1/3)jk+ are draws from the Dirichlet distribution, with

p(πjk+|π(t)j+l, π

(t)+kl, Y

(t)) ∝J∏j=1

K∏k=1

παjk++µ(t)jk+−1, (3.13)

and g+++ =∑

JK gjk+. gj+L and g+kl are drawn subsequently with their correspond-

ing restrictions. Similarly to the ECM algorithm (Section 2.6) a single run through

the CP steps each iteration is sufficient. This helps to speed up convergence.

More details on both of these algorithms and practical advice on implementation

is available in chapter 4 of Schafer [1997].

3.4 Extensions to the DA algorithm

As outlined in Section 2.1, there are many novel issues found in the meta-analysis

datasets we have analyzed. While data augmentation methods have been researched

for the general case of multiple partially classified tables, extensions are required

in order to deal with these extra complications. Here we shall outline each of the

problems and the solution we have developed.

(i) We may observe all the variables in Oi, but at fixed levels of each of the variables

in Ci (a slice). Here Oi ∪ Ci = F is the model term contributed by the study to

the Bayesian IPF step and a modification to the I-step of the DA algorithm is

36 CHAPTER 3. DATA AUGMENTATION

necessary. If ci are the actual levels of the variables in Ci that are observed, our

observed partial table can be written niriOi|Ci=ci . It should be noted that this

section of the contingency table is in fact fully classified, with the remainder

of the table missing, i.e. Ci 6= ci or C ′i . Therefore we need only generate

multinomial samples in the section Oi|C′i with the distribution:

M

(ni

∑πOi|C

′i∑

πOi|Ci, πOi|C

′i

), (3.14)

where πOi|C′i

=∑C′iπF . This generated table collated with nir

iOi|Ci=ci will con-

stitute the output from the imputation step.

(ii) If we observe a slice in some variables, while some factors are also missing

(marginal), then our strategy is somewhat similar. The model term corresponds

to Oi ∪ Ci as these are the only terms observed in the study. The two sections

of the table, the slice and the non-slice, may be generated separately as they

contain variable sets which are disjoint. For the slice section (Oi ∪Ci ∪Mi), we

generate from the product multinomial distribution

M(niri(Oi∪Ci), πMi|(Oi∪Ci)). (3.15)

There is a two step process for the non-slice section (Oi ∪ C′i ∪Mi). Firstly we

deal with sample across the non-observed conditional levels to find A,

A ∼M

(ni

∑π(Oi∪Mi)|C

′i∑

π(Oi∪Mi)|Ci, πOi|(Mi∪C

′i)

)(3.16)

We then expand this multiway table over the missing margin, via a product

multinomial distribution once more,

M(A, πMi|(Oi∪C′i)). (3.17)

The distributions resulting from the steps in equations 3.15 and 3.17 are then

collated to provide the output of the I-step.

3.5. SIMULATION STUDIES 37

(iii) We may observe multiple (J) slices in a single study, each with its associated

set of missing variables. Again, adjustments are required to both the I-step

and the model term. There are in fact a collection of model terms relating to

this study,O1i ∪ C1

i , . . . ,OJi ∪ CJi

. For each of these slices we carry out the

algorithm as it is outlined in (ii) above, generating a separate full model for

each. The sum of these J fully classified tables provides the input from study

i for the P-step of the algorithm. The methods outlined here assume that the

slices are in fact disjoint. If they are not, the I-step may become a complex task

involving factored posterior generation. We did not encounter such difficulties

in the data sets in this dissertation.

(iv) The retrospective sampling adjustment is relatively straightforward to imple-

ment in the data augmentation framework when using Bayesian IPF. We put a

final constraint on the IPF to ensure that the case-control totals are equivalent

to the those of the population for each of the constituent studies. This results

in an extra CP step for each observed study and is similar in nature to the pro-

posed likelihood-based approach. There are more details on the use of Bayesian

models for retrospective sampling schemes available in Seaman and Richardson

[2001], however the multiple study version has not been previously dealt with

elsewhere.

3.5 Simulation studies

To establish the equivalency between the likelihood-based approach and the Bayesian

methods outlined in this chapter, a simple simulation study was carried out. A four-

way table was formed, with each of its factor containing three levels, hence a 3x3x3x3

contingency table. The details of the simulation were as follows:

• 10 random Poisson samples, constituting the observed tables.

• Each observed table had a sample size of 200.

38 CHAPTER 3. DATA AUGMENTATION

• Each table had a random missingness patterns constructed so as to mimic those

outlined in Chapter 2, i.e. each observed table/sample contained missing and

conditional variables, with no single complete table observed.

The EM and DA algorithms are applied to this simulated data and the parameter

estimates found under the respective models are shown in Figure 3.1. Here we also

compare against the true underlying factor relationships, from the original generating

table. The four marginal tables are provided, with the EM and DA algorithms both

providing accurate estimates. This equality in performance between the two models

was also replicated in further simulation studies performed.

Figure 3.1: Simulation results comparing the marginal parameter estimates under thelikelihood-based approach and the Bayesian methods

Chapter 4

Variance Estimation

4.1 The sandwich estimate

In this section we will consider robust parameter estimation. It is reasonable to

say that the model we choose to fit is often not the true underlying probability

structure which generated the data. While this seems at first glance to be detrimental,

here we will outline methods developed which correct for this model misspecification.

Some features of the distribution can still be consistently estimated and it is possible

to produce unbiased variance estimates. In particular we will concentrate on how

maximum likelihood estimation performs under such conditions.

Suppose x1, x2, . . . , xn are an iid sample from an unknown distribution g(x) and

our model is fθ(x). Maximizing the likelihood is equivalent to

maximizing1

n

∑i

log fθ(xi), (4.1)

which in large samples is equivalent to,

minimizing − Eg log fθ(x) = −∫g(x) log fθ(x)dx. (4.2)

Since Eg log g(x) is an unknown constant wrt θ it is also equivalent to minimizing the

39

40 CHAPTER 4. VARIANCE ESTIMATION

Kullback-Leibler distance,

minimizing D(f, g) = Eg log g(x)− Eg log fθ(x) (4.3)

Hence maximizing the likelihood is equivalent to finding the distribution closest to

the truth under the Kullback-Leibler distance measure.

In truth x1, x2, . . . , xn are an iid sample from an unknown distribution g(x), while

we assume a model fθ(x). θ is the maximum likelihood estimate based on this assumed

model. θ0 is the parameter being estimated by the ML procedure and is the maximum

of λ(θ) ≡ Ex log fθ(x). We expect that θp→ θ0 (Pawitan [2001] Theorem 13.1).

Let θ be a consistent estimate of θ0, assuming the model fθ(x) . Allow θ to be a

vector and define

J = E

(∂fθ(x)

∂θ

)(∂fθ(x)

∂θ′

)|θ=θ0 (4.4)

I = −E(∂2fθ(x)

∂θ∂θ′

)|θ=θ0 (4.5)

J and I are identical if fθ0(x) is the true model, and hence in this case the estimated

variance is the “naive” inverse Fisher information.

Theorem A. Assuming the standard regularity conditions 1√n(θ−θ0)

d→ N(0, I−1J I−1)

Proof: The log-likelihood of θ is

logL(θ) =∑i

log fθ(xi) (4.6)

Using a Taylor series approximation, we expand the score function around θ,

logL(θ0)

∂θ=

∂ logL(θ0)

∂θ|θ=θ +

∂2 logL(θ∗)

∂θ∂θ′(θ − θ) (4.7)

=∂2 logL(θ∗)

∂θ∂θ′(θ − θ) (4.8)

4.1. THE SANDWICH ESTIMATE 41

where |θ∗ − θ| ≤ |θ − θ| and let

yi ≡∂ log fθ(xi)

∂θ. (4.9)

Therefore,∂ logL(θ)

∂θ=∂∑

i log fθ(xi)

∂θ=∑i

∂ log fθ(xi)

∂θ=∑i

yi (4.10)

and so∂ logL(θ)

∂θis the sum of iid yi, with mean

E(Yi) = E∂ log fθ(xi)

∂θ(4.11)

=∂E log fθ(xi)

∂θ= λ

′(θ) (4.12)

At θ = θ0, EYi = 0 and variance

var(Yi) = J = E

(∂fθ(x)

∂θ

)(∂fθ(x)

∂θ′

)|θ=θ0 (4.13)

By the central limit theorem at θ = θ0

1Regularity conditions [Lehmann and Casella, 1998]:

(a) The parameter space Ω is an open interval (not necessarily finite).

(b) The distribution Pθ of the Xi have common support, so that the set A = x : fθ(x) > 0 isindependent of θ.

(c) For every x ∈ A, the density fθ(x) is twice differentiable under w.r.t. θ, and the secondderivative is continuous in θ.

(d) The integral∫fθ(x)dµ(x) can be twice diffentiated under the integral sign.

(e) The Fisher information I(θ) satistfies 0 < I(θ) <∞.

(f) For any given θ0 ∈ Ω, there exists a positive number c and a function M(x) (both of which

may depend on θ0) such that∣∣∣∣∂2 log fθ(x)

∂θ2

∣∣∣∣ ≤ M(x), ∀x ∈ A, θ0 − c < θ < θ0 + c and

Eθ0 [M(x)] <∞.

(g) E

[∂ log fθ(x)

∂θ

]= 0.

(h) E

[−∂

2 log fθ(x)∂θ2

]= E

[∂ log fθ(x)

∂θ

]2= I(θ).

42 CHAPTER 4. VARIANCE ESTIMATION

1√n

∑i

Yid→ N(0,J ). (4.14)

Since θp→ θ0,

1

n

∂2 logL(θ∗)

∂θ∂θ′=

1

n

∑i

∂2 log fθ∗(xi)

∂θ∂θ′(4.15)

p→ E∂2 log fθ∗(X)

∂θ∂θ′|θ=θ0 = −I (4.16)

Therefore,

1√n

∂ logL(θ0)

∂θ=

1

n

∂2 logL(θ∗)

∂θ∂θ′√n(θ0 − θ). (4.17)

By Slutsky’s theorem,

√n(θ − θ0)

d→ N(0, I−1J I−1) (4.18)

Hence we can find the estimated variance at the MLE as being

I−1(θ)J (θ)I−1(θ). (4.19)

The estimate I−1(θ) in the equation above is computed as part of the Newton-

Raphson algorithm for ML estimation and J (θ) is found as a byproduct of the scoring

algorithm. However, things are not simple in the case of missing data, where the EM

algorithm is used in finding the MLE’s. The next few sections of this chapter will

outline the complications involved in reproducing an accurate sandwich estimate in

such situations, and outline some of proposed solutions.

4.2. EXTENDING THE SANDWICH ESTIMATE TO MISSING DATA 43

4.2 Extending the sandwich estimate to missing

data

As outlined in the previous section, the constituent parts of the large-sample covari-

ance matrix 4.19 are not produced as direct by-products of the EM algorithm and

hence must be produced independently of this process. The observed information

matrix I−1(θ) is in fact stated more correctly as I−1( ˆθ|Yobs) in this setting, as we

have both observed and unobserved data. This can be found directly as the second

derivative of the log-likelihood, calculated at θ = θ. Unfortunately this work may be

restricted by computational restrictions, especially in the inversion of the information

matrix.

Alternatively, one may calculate the information matrix as the difference of the

complete data information and the missing information [Meng and Rubin, 1991]:

I(θ|Yobs) = −∂2Q(θ|θ)∂θ∂θ′

+∂2H(θ|θ)∂θ∂θ′

(4.20)

where,

Q(θ|θ(t)) =

∫[`(θ|Yobs, Ymis)] f(Ymis|Yobs, θ(t))dYmis (4.21)

H(θ|θ(t)) =

∫[log f(Ymis|Yobs, θ)] f(Ymis|Yobs, θ(t))dYmis (4.22)

Due to the missingness in the data, numerical approximations are applied only in

calculating the missing information (matrix), the term involving H(θ|θ). Hence, this

can be lead to unstable estimates of the covariance matrix.

There are a collection of large-sample methods developed from the sandwich esti-

mator. Foremost amongst them is the SEM algorithm, outlined in the next section.

Other noteworthy alternatives include those developed by Louis [1982], which requires

the calculation of the conditional expectation of the squared complete-data score func-

tion, and Little and Rubin [2002] involving a two part quadratic approximation to

the likelihood.

44 CHAPTER 4. VARIANCE ESTIMATION

4.3 Supplemented EM

Supplemented EM [Meng and Rubin, 1991] was introduced as an alternative method

to estimate the variance-covariance matrix, specifically when the EM algorithm is

used to find parameter estimates. Advantages for this method include:

1 Uses only code from the E and M steps

2 Does not require the missing information explicitly, uses only the large-sample

complete data variance-covariance matrix (Vc)

3 Only standard matrix operations required

4 More stable than numerically differentiating l(θ|Yobs)

Using the notation in Meng and Rubin [1991], we define

icom =∂2Q(θ|θ)∂θ∂θ′

|θ=θ∗

imis =∂2H(θ|θ)∂θ∂θ′

|θ=θ∗

iobs = I(θ|Yobs)|θ=θ∗ (4.23)

DM is the derivative of the EM mapping (M). Even though M does not have an

explicit mathematical form, its derivative DM can be estimated from the output

of forced EM steps, whereby we effectively numerically differentiate M . This is the

central concept of the SEM algorithm. So in effect DM represents the fraction of

missing information in the gradient of the EM mapping, and hence controls the speed

of convergence with,

DM = imisi−1com = I − iobsi−1

com (4.24)

We denote the converged value of θ to be θ∗. Therefore,

4.3. SUPPLEMENTED EM 45

Vobs = i−1obs

= Vcom(I −DM)−1

= Vcom(I −DM +DM)(I −DM)−1

= Vcom + VcomDM(I −DM)−1

= Vcom + ∆V (4.25)

Firstly we obtain the MLE θ and then a sequence of SEM iterations are run,

iteration (t+ 1) taking the following form:

Input: θ and θ(t).

Step 1: Run the usual E and M steps to obtain θ(t+1).

Step 2: Fix i = 1 and calculate

θ(t)(i) = (θ1, . . . , θi−1, θ(t)i , θi+1, . . . , θd) (4.26)

which is θ, with the ith component replaced by θ(t)i .

Step 3: Treating θ(t)(i) as the current estimate of θ , run one iteration of EM

to obtain θ(t+1)(i).

Step 4: Find the ratio,

r(t)ij =

θ(t+1)(i)− θjθ

(t)i − θi

, for j = 1, . . . , d (4.27)

Output: θ(t+1) andr

(t)ij : i, j = 1, . . . , d

.

Hence we can find DM as limt→∞ rij, with the element rij found once the sequence

r(t∗)ij , r

(t∗+1)ij , . . . is stable for some t∗. It is likely that different values of t∗ will be used

for different elements rij. When all elements in the ith row of DM have been obtained,

steps 2-4 are no longer required for that value of i in subsequent iterations.

46 CHAPTER 4. VARIANCE ESTIMATION

Once we have found the converged DM, it is easy to reconstruct the observed

variance Vobs, using Equation 4.25. This method is designed so as to estimate the

variance of the unknown parameters θ, but the variances of other quantities of interest

(e.g. cell probabilities) may be reconstructed once the full covariance function for θ

has been established.

In practice we did find some limitations in the SEM algorithm:

• Despite the claims in the paper, this method does not necessarily produce a sym-

metric covariance matrix, with large discrepancies noted in large dimensional

cases.

• DM is often not positive semi-definite, small adjustments to the eigenvalues are

required to correct for this.

• Each parameter converges at a different rate, so it requires quite a bit of hand-

tuning in order to confirm convergence across all parameters.

• Tends to overestimate the variance slightly, a drawback outlined in the original

paper also.

4.4 The jackknife

The jackknife method is employed in order to estimate the standard error of the

prediction from the fitted model. A version of cross-validation, the jackknife uses

the leave-one-out method to estimate the bias and the standard error of an estimate.

Given the full data, x = (x1, ..., xn), the jackknife methods creates n sub-samples,

leaving one sample out each time: x(i) = (x1, ..., xi−1, xi+1, ..., xn) for i = 1, . . . , n.

Therefore the parameter of interest θ is estimated in each sub-sample by θ(i) = f(x(i)),

with θ(i) the jackknife replication of θ. The jackknife estimate of bias is defined as

biasjack = (n− 1)(θ(.) − θ) (4.28)

4.5. THE BOOTSTRAP 47

where

θ(.) =1

n

n∑i=1

θ(i). (4.29)

The jackknife estimate of the standard error is defined as

sejack =

[n− 1

n

n∑i=1

(θ(i) − θ(.)

)2] 1

2

. (4.30)

This method can be adapted to deal with data such as those we have encountered

in this dissertation. Rather than excluding a single sample, we omit one full study

at a time, producing a set of S parameter estimates θ(1), . . . , θ(S). There are also

some beneficial side-effects in using the jackknife variance method. The jackknife

samples may be used to cross-validate for model-selection (Section 2.10). As explained

in Section 2.8 faulty data or aberrant studies can be identified using the jackknife

influence.

4.5 The bootstrap

The jackknife may be regarded as a special case of the bootstrap [Efron and Tibshi-

rani, 1993], in the general family of resampling techniques. The general bootstrap

method provides a computer-based nonparametric estimate of the standard error, but

with similar asymptotic properties to the sandwich estimator. Each bootstrap sam-

ple is a sample with replacement of size n from the observations, and B independent

samples such as this are found, x∗1, . . . ,x∗B. θ∗(b) is the ML estimate of θ based on

the bth bootstrap sample x∗b. Therefore the overall bootstrap estimate of θ is

θboot =1

B

B∑b=1

θ(b), (4.31)

and the bootstrap estimate of the standard error of θ or θboot is

seboot =

[1

B − 1

B∑b=1

(θ(b) − θboot)2

]1/2

(4.32)

48 CHAPTER 4. VARIANCE ESTIMATION

As in the jackknife, samples are formed here on a study-to-study basis in the meta-

analysis setting; choosing n studies (from the total of S studies) for each bootstrap

sample.

The bootstrap yields valid large-sample standard errors regardless of the validity of

the assumptions of the underlying model. It is however limited in cases with moderate

data sets or extensive missing data, as many of the bootstrap samples will exclude

the data required for vital parameter estimates. Therefore the jackknife method is

more appropriate for fitting the sparsely classified multiway tables observed in this

thesis, as the resampling excludes a smaller proportion of the pivotal data.

4.6 Bayesian posterior

The data augmentation algorithm outlined in Chapter 3 produces an accurate esti-

mate of variance when a uniform prior is used. Under this prior the posterior mode is

the ML estimate and the posterior variance is a consistent estimate of the large-sample

variance of the ML estimate. It is an attractive approach due to this asymptotic prop-

erty, in addition to its superior performance in small sample cases. Here it provides

inference based directly on the posterior distribution without invoking large-sample

normal approximations.

We denote θ(1), . . . , θ(M) as the resulting estimates from a single simulation run of

the data augmentation algorithm of length M . We shall discard the output during

the initial burn-in period. For any scalar function φ, we define φ(t) = φ(θ(t)) and its

sample average as

φ =1

M

M∑t=1

φ(t). (4.33)

The sample variance of φ overestimates the true variance V (φ), because the elements

in the sequence φ(1), . . . , φ(M) are correlated. This is because the input for each

iteration of the data augmentation algorithm is the output from the last. The single

run sample variance does at least provide a crude lower bound for the variance.

Therefore, it is essential to adjust for this autocorrelation. This may be achieved

4.6. BAYESIAN POSTERIOR 49

via subsampling; averaging over every bth iterate instead:

φ(b) =1

m

m∑t=1

φ(tb), (4.34)

where m = M/b. If the choice of b is made carefully (and checked by inspection of

the sample ACF), then we may estimate V (φ) by:

V (φ(b)) =1

m

m∑t=1

(φ(tb) − φ(b))1/2. (4.35)

Unfortunately this method tends to overestimate the true variance V (φ) [Schafer,

1997].

A more stable and reliable method for posterior variance estimation is found using

multiple chain simulation of the algorithm. We shall perform R replicate runs from

a common starting distribution, again discarding data from the burn-in period. A

sample of size M is then found from each run, with the tth estimate of θ from the rth

run denoted θ(r:t). The within-run sample average is

φ(r) =1

M

M∑t=1

φ(r:t), (4.36)

and the pooled sample average is found as

φ =1

RM

r=1∑R

M∑t=1

φ(r:t). (4.37)

Hence we may find an unbiased estimate of the variance of a single φ(r) using the

between-run variance

B =1

R− 1

r=1∑R

(φ(r)−)1/2 − φ, (4.38)

withB

Ra reasonable approximation to the variance of the pooled estimate, V (φ).

50 CHAPTER 4. VARIANCE ESTIMATION

4.7 Simulation studies

Figure 4.1: Variance estimation: simulation results for cells 1 to 4

A simulation study was carried out to compare the performance of these variance

estimate techniques using data similar to that outlined in Chapter 2, with the details

as follows:

• 300 simulations were run.

• Each simulation contained 10 random Poisson samples, constituting the ob-

served tables.

• Each observed table had a sample size of 30.

4.7. SIMULATION STUDIES 51

• The true underlying ”full” contingency table had three factors, each with two

levels (2x2x2 table). A table of this low-order in magnitude was necessary so

as to allow for feasible computation. Since there are eight cells in total, this is

what is observed in the associated plots.

• Each table had a random missingness patterns constructed so as to mimic those

outlined in Chapter 2, i.e. each observed table/sample contained missing and

conditional variables, with no single complete table observed.

We generated random Poisson counts from the true underlying distribution, and

compared the estimates produced by the candidate methods against the true standard

deviation, shown in red in each of the plots (calculated using the multinomial variance

formula). Each of the other five lines shows the cumulative estimate of the standard

deviation for its respective method, after that number of simulation. Therefore a

suitable method should converge to the truth as the number of simulations increases.

Figures 4.1 and 4.2 show the results of this analysis for each of the eight cells in the

table. We have also reproduced the results from a single cell in Figure 4.3 in order to

provide a clearer view of the plot details for at least one element.

It is clear from this simulation and others we also completed, that the SEM is not a

viable method for this analysis. It overestimates the variance considerably in many of

these plots and required quite a high level of parameter-tuning by the user to achieve

results even this accurate. The data augmentation methods were quite accurate and

tended to follow-one another closely. This is perhaps due in part to a common starting

distribution and random seed. The bootstrap results were not shown in these plots, as

they were quite similar to those of the jackknife. The jackknife method was perhaps

the most accurate of all the methods here, and this was replicated in other simulation

studies also. Combined with the additional side-benefits as outlined above in Section

4.4, the authors conclude that this is the most suitable variance estimation method

for the likelihood-based models.

52 CHAPTER 4. VARIANCE ESTIMATION

Figure 4.2: Variance estimation: simulation results for cells 5 to 8

4.7. SIMULATION STUDIES 53

Figure 4.3: Investigating alternatives, one sample

Chapter 5

Retrospective Adjustment

5.1 Description of the problem

Meta-analysis attempts to aggregate the results of many studies in order to leverage

their combined power. It is therefore often necessary to combine effect size estimates

from numerous constituent studies. In this chapter we will address the issues in-

volved in combining fixed effect or random effect estimates for categorical data, and

in particular we will discuss the issues involved in adjusting for retrospective sampling.

For example and without loss of generalization we shall look at S studies, each

providing a 2x2 table for a set of factors A and B. The number of subjects in study s

who observed the i, j combination of factors A and B is denoted n(s)ij . A fully observed

set of tables may be denoted:

B1 B2

A1 n(1)11 n

(1)12

A2 n(1)21 n

(1)22

B1 B2

A1 n(2)11 n

(2)12

A2 n(2)21 n

(2)22

. . . . . .

B1 B2

A1 n(S)11 n

(S)12

A2 n(S)21 n

(S)22

Table 5.1: Combining S fully observed two-way tables

In sections 5.2, 5.3 and 5.4 we will introduce contemporary methods for combin-

ing such tables for prospective studies. While in sections 5.5, 5.6, 5.7 and 5.8 we

will develop new techniques for adjusting these methods to account for retrospective

sampling schemes.

54

5.2. MAXIMUM LIKELIHOOD METHOD 55

5.2 Maximum likelihood method

The maximum likelihood method is most easily implemented via logistic regression.

We treat the studies as a third variable C and we now use the notation πjs =n

(s)1j

n(s)1.

=

P (A = 1|B = j, C = s). Using the model

logit(πjs) = α + βxj, j = 1, 2. (5.1)

where x1 = 1, x2 = 0. This model assumes that the AB conditional odds ratio is

the same at each category of C, namely exp(β). The maximum likelihood estimate

of the common odds ratio is exp(β)

The ML estimate β of the log odds ratio tends to be too large in absolute value

when S is large and the data are sparse. For example in the sparse-data case, with

only a single matched pair in each study, βp→ 2β as n→∞.

5.3 Mantel-Haenszel method

Mantel and Haenszel [1959] proposed a non model-based approach to this problem.

Here the joint odds ratio estimate oMH is

oMH =

∑Ss=1 n

(s)11 n

(s)22 /n

(s)..∑S

s=1 n(s)12 n

(s)21 /n

(s)..

. (5.2)

Robins et al. [1986] derived an associated non-null standard error estimate for

log(oMH):

SE =1∑A(s)√

2

(∑A(s)B(s) + oMH

∑(B(s)C(s) + A(s)D(s)

)+ o2

MH

∑C(s)D(s)

) 12

(5.3)

where,

56 CHAPTER 5. RETROSPECTIVE ADJUSTMENT

A(s) =n

(s)11 n

(s)22

n(s)..

(5.4)

B(s) =n

(s)11 + n

(s)22

n(s)..

(5.5)

C(s) =n

(s)12 n

(s)21

n(s)..

(5.6)

D(s) =n

(s)12 + n

(s)21

n(s)..

(5.7)

5.4 Pooling log-odds ratios

A widely used alternative to the Mantel-Haenszel techniques is the pooling of the log

odds ratios across the S studies. An adjustment factor of 0.5 is made to each cell

count to avoid a divide by zero error. Define:

Ls = ln(n

(s)11 + 0.5)(n

(s)22 + 0.5)

(n(s)12 + 0.5)(n

(s)21 + 0.5)

, (5.8)

and

Ws =

(∑i

∑j

1

n(s)ij + 0.5

)−1

. (5.9)

The pooled log odds ratio is

L* =

∑WsLs

(∑Ws)1/2

, (5.10)

with a standard error of

SE =1

(∑Ws)1/2

. (5.11)

When the cell frequencies within all of the S studies are large, the Mantel-Haenszel

estimate will be close in value to the estimate obtained by pooling the log odds ratio.

There is disagreement however in cases where the cell frequencies are small. In such

circumstances the Mantel-Haenszel estimator is superior to the log odds ratio [Hauck,

5.5. MODIFICATION FOR RETROSPECTIVE SAMPLING 57

1979]. There have been other alternatives proposed, but these competitors have been

shown to be inferior in the case of stratified studies [Agresti, 2002].

5.5 Modification for retrospective sampling

The methods proposed thus far have dealt with data collected from prospective re-

search. However, in retrospective sampling designs, such as case-control biomedical

studies, an adjustment is required in order to correctly estimate the effects. This is

because in case-control studies the explanatory variable X is random, rather than the

response variable Y. Anderson and Richardson [1979] proposed a solution to this prob-

lem for the single study case, using Bayes theorem and the logit link function. Let Z

indicate whether each subject is sampled (1=yes, 0=no), with ρ0 = P (Z = 1|Y = 1)

denoting the probability of sampling a case and ρ1 = P (Z = 1|Y = 0) denoting the

probability of sampling a control.

P (Y = 1|Z = 1, X = x) =P (Z = 1|Y = 1, X = x)P (Y = 1|X = x)∑1j=0 P (Z = 1|Y = j,X = x)P (Y = j|X = x)

=P (Z = 1|Y = 1)P (Y = 1|X = x)∑1j=0 P (Z = 1|Y = j)P (Y = j|X = x)

=ρ1exp(α + βx)

ρ0 + ρ1exp(α + βx)

=exp(α + log(ρ1/ρ0) + βx)

1 + exp(α + log(ρ1/ρ0) + βx)(5.12)

So by fitting a logistic model, the estimated effect parameter β is equivalent to that

produced by a prospective study. There is an intercept parameter change however

with

α∗ = α + log(ρ1

ρ0

) (5.13)

Hence, when attempting prediction in a case-control study it is necessary to adjust

only the intercept parameter for the fitted logistic model. This can be done by

adjusting the estimates on the logit scale, using external information about the actual

disease prevalence in the entire population.

58 CHAPTER 5. RETROSPECTIVE ADJUSTMENT

Since,

ρ1 = P (Z = 1|Y = 1) =P (Y = 1|Z = 1)P (Z = 1)

P (Y = 1)

ρ0 = P (Z = 1|Y = 0) =P (Y = 0|Z = 1)P (Z = 1)

P (Y = 0). (5.14)

Therefore,

log

(ρ1

ρ0

)= log

(P (Y = 1|Z = 1)

P (Y = 0|Z = 1)

)− log

(P (Y = 1)

P (Y = 0)

), (5.15)

with P (Y = 1|Z = 1) = P (Case|Sampled) and P (Y = 1) = P (Case) in entire

population.

5.6 Extension of the modification for retrospective

studies

However, in the case of multiple studies investigating the effect size of the same factor

it is necessary to extend the model outlined above. This is because the case-control

mix in each of the constituent studies will be different. Here

ρ(s)1 = P (Z = 1|Y = 1, S = s) (5.16)

denotes the probability of sampling a case in the sth study, with

ρ(s)0 = P (Z = 1|Y = 0, S = s) (5.17)

5.6. EXTENSION OF THE MODIFICATION FOR RETROSPECTIVE STUDIES59

the corresponding probability of sampling a control.

P (Y = 1|Z = 1, X = x, S = s) =P (Z = 1|Y = 1, X = x, S = s)P (Y = 1|X = x, S = s)∑1j=0 P (Z = 1|Y = j,X = x, S = s)P (Y = j|X = x, S = s)

=P (Z = 1|Y = 1, S = s)P (Y = 1|X = x, S = s)∑1j=0 P (Z = 1|Y = j, S = s)P (Y = j|X = x, S = s)

(s)1 exp(αs + βx)

ρ(s)0 + ρ

(s)1 exp(αs + βx)

=exp(αs + log(ρ

(s)1 /ρ

(s)0 ) + βx)

1 + exp(αs + log(ρ(s)1 /ρ

(s)0 ) + βx)

(5.18)

The only assumption made in this section is that

P (Z = 1|Y = 1, X = x, S = s) = P (Z = 1|Y = 1, S = s) (5.19)

i.e. the sampling probabilities do not depend on the covariate of interest. Again we

have

ρ(s)1 = P (Z = 1|Y = 1, S = s) =

P (Y = 1|Z = 1, S = s)P (Z = 1)

P (Y = 1)

ρ(s)0 = P (Z = 1|Y = 0, S = s) =

P (Y = 0|Z = 1, S = s)P (Z = 1)

P (Y = 0)(5.20)

and the adjustment has the same form,

α∗s = αs + log

(s)1

ρ(s)0

), (5.21)

with,

log

(s)1

ρ(s)0

)= log

(P (Y = 1|Z = 1, S = s)

P (Y = 0|Z = 1, S = s)

)− log

(P (Y = 1)

P (Y = 0)

)(5.22)

The full logistic model is

logit(πjs) = αs + βxj, j = 1, 2. (5.23)

60 CHAPTER 5. RETROSPECTIVE ADJUSTMENT

So this adjustment must be made separately for the adjusted cell predictions from

each of the S studies. The exp(β) is the maximum likelihood estimate of the odds

ratio exp(β). This method can be extended easily to case where X has more than

two levels or is a continuous variable, and even to cases with multiple parameter

estimation.

5.7 Loglinear-logit model connection

Logit models consider a single categorical response variable and it’s relationship with

a group of explanatory variables. Loglinear models, by contrast, treat all categori-

cal equivalently, focusing on associations and interactions in their joint distribution.

There is however, a well-established equivalency between loglinear and logistic mod-

els when the categorical (response for logit) variable, Y, is binary. For example if we

consider a three-way table and the fitted loglinear model (XY,XZ,YZ), the logit of Y

is:

logP (Y = 1|X = i, Z = k))

P (Y = 1|X = i, Z = k))= log

µi1kµi0k

= log (µi1k)− log (µi0k)

=(λ+ λXi + λY1 + λZk + λXYi1 + λXZik + λY Z1k

)−(λ+ λXi + λY0 + λZk + λXYi0 + λXZik + λY Z0k

)=

(λY1 − λY0

)+(λXYi1 − λXYi0

)+(λY Z1k − λY Z0k

)(5.24)

Therefore, the logit has the additive form:

logit [P (Y = 1|X = i, Z = k] = α + βXi + βZk (5.25)

whereby the first parameter is a constant, the second parenthetical term depends

only on the category i of X and the third parameter depends solely on category k of

Z. This may be denoted an (X + Z) logistic model. Table 5.2 outlines some more

equivalent models in the three-way example.

5.8. MODIFICATION IN THE LOGLINEAR SETTING 61

Loglinear Model Logistic Model(Y,XZ) (-)

(XY,XZ) (X)(YZ,XZ) (Z)

(XY,YZ,XZ) (X + Z)(XYZ) (XZ)

Table 5.2: Equivalent loglinear and logistic models for a three-way contingency tablewith a binary response variable Y

5.8 Modification in the loglinear setting

As previously outlined in sections 5.5 and 5.6, retrospective sampling scheme param-

eter estimates may be adjusted in order to produce accurate probability estimates.

All of this work has been outlined in the logistic model setting. However, in loglinear

models we can use a similar adjustment scheme. In logistic terms, in the three-way

example with (XZ), we have:

logit [P (Y = 1|X = i, Z = k)] =

[α + log

ρ1

ρ0

]+ βXi + βZk + βXZik (5.26)

Therefore to provide equivalency in the loglinear model (XYZ) it is required that:

λ∗Y1 − λ∗Y0 = α + logρ1

ρ0

(5.27)

where previously,

λY1 − λY0 + logρ1

ρ0

= α (5.28)

⇒(λ∗Y1 − logρ0

)−(λ∗Y0 − logρ1

)= α = λY1 − λY0 (5.29)

62 CHAPTER 5. RETROSPECTIVE ADJUSTMENT

Thus the intercepts for cases (Y = 1) and the controls (Y = 0) are adjusted separately

using,

ρ1 =P (Y = 1|Z = 1)

P (Y = 1)(5.30)

ρ0 =P (Y = 0|Z = 1)

P (Y = 0)(5.31)

with Z = 1 for sampled observations. This method is extended as we did in section

5.6 to deal with multiple studies, with each study having it own set of intercept

adjustment factors. In terms of practical model fitting, this is implemented as an

offset term.

In this chapter thus far we have not mentioned how to deal with retrospective

sampling in cases where there are slices of information from some studies. The general

procedure mentioned in the Section 2.5 referred to an E-step in such cases where the

estimated table is modified to include such slices. In these cases only the full expanded

table riF must be adjusted as above.

5.9 Simulation studies

In order to confirm the findings in this chapter, we carried out simulation studies with

both retrospective and prospective sampling schemes. The details of these sampling

schemes were as follows:

• Random multinomial draws were made from a five-way contingency table, with

the five factors having two, two, two, three and four levels respectively (2x2x2x3x4

table).

• Each table had a random missingness patterns constructed so as to mimic those

outlined in Chapter 2, i.e. each observed table/sample contained missing and

conditional variables, with no single complete table observed.

• Two sets of data were produced in each simulation run:

5.9. SIMULATION STUDIES 63

1. The first of these had a preassigned (random) reweighting factor, which is

consistent with retrospective studies.

2. The second did not have this reweighting factor and hence simulated the

more straightforward prospective studies.

• Two loglinear models were fit, with the offsets for the case-control data.

• We then analyzed the performance of the case-control adjustment developed in

this chapter.

In the first set of plots, Figure 5.1, 10 random studies were generated for each data

point. We looked over a range of sample sizes per study to see if this affected the

accuracy of the fitted models. Unsurprisingly, large sample sizes resulted in a better

fit. A similar relationship between model fit and the number of studies would be

expected and this is witnessed in Figure 5.2 (with sample size fixed at 40 per study). In

both figures, the top two plots consider the overall fit via Kullback-Leibler divergence

and Euclidean distance, while the other two plots investigate prediction error for

two particular cells in the contingency table. It is apparent that the adjustment for

retrospective sampling has succeeded, as the performances of the retrospective and

prospective models are shown to be similar in this analysis.

An issue not discussed thus far is the estimate of the population disease incidence

rate for the , and in particular the effect of a misspecified rate. Of course this is a

difficult thing to estimate for some diseases, but small errors in this estimation can

lead to large errors in the model predictions as seen in Figure 5.3. In this simulation

study the true population disease rate (P (Y = 1) = P (Case)) is 0.2 and we have

generated random multinomial samples at this rate. The assumed rate used in the

retrospective adjustment was varied in the range from 0 to 1, and we can see that

a small misspecification in this rate can lead to a large divergence in the resulting

model from the true underlying distribution.

64 CHAPTER 5. RETROSPECTIVE ADJUSTMENT

Figure 5.1: Confirming the retrospective adjustment for loglinear models with varyingsample size, as sample size increases both the retrospective and prospective modelsprovide similarly better estimates

5.9. SIMULATION STUDIES 65

Figure 5.2: Confirming the retrospective adjustment for loglinear models with varyingthe number of studies.

66 CHAPTER 5. RETROSPECTIVE ADJUSTMENT

Figure 5.3: Misspecification of the population disease rate

Chapter 6

Psoriasis Meta-Analysis

6.1 Psoriasis

Psoriasis is a chronic non-contagious autoimmune disease affecting the skins and

joints of patients. Often the symptoms include thick red scaly patches on the skin or

fingernails, the severity of which can range from small localized patches to full body

coverage. It has been acknowledged in medical literature as early as ancient Greece,

where psora was used to describe an itchy skin condition, and as tzaraat in the Bible.

However, the condition name psoriasis was not termed until 1841 by Ferdinand von

Hebra, a Vienese dermatologist [Meenan, 1955]. It is estimated that approximately

2% of the population worldwide are affected by psoriasis [Griffiths and Barker, 2007],

with 35% of patients classified as having moderate to severe symptoms (> 3% of body

surface). Both males and females suffer from this disease and it may occur at any age,

although the majority of cases are initially diagnosed between the ages of 15 and 25

years. Diagnosis of the disease is made based on the appearance of the the skin alone,

to date no special blood tests or diagnostic procedures have been developed. There

are many treatments available, but because of its chronic recurrent nature psoriasis

is a challenge to treat.

There are in fact two main types of psoriasis which we will deal with it in this

analysis:

1. Type I: Also known as early onset psoriasis, these are cases where chronic plaque

67

68 CHAPTER 6. PSORIASIS META-ANALYSIS

first appears before 40 years. Approximately 75% of patients suffer from this

type of psoriasis.

2. Type II: For late onset psoriasis, symptoms present after 40 years. Although

this is the rarer form of the disease, it is the potentially fatal variant.

While much investigation into the cause of psoriasis has been carried out, the

mechanism is still not fully understood. There are two main hypotheses about the

process leading to the development of the disease. The first such theory states that

the disease can be simply traced to faults in the epidermis and its keratinocytes. The

second hypothesis views the disorder as being an immune-mediated disease, where

the skin inflammations are secondary to factors produced by the immune system. It

states that the excessive production of skin cells is in fact initiated by the activation

of T cells and their migration to the dermis. Unfortunately, we do not yet understand

why the T cells become activated, as their natural function is to help protect the body

against infection.

Psoriasis remains an idiosyncratic disease; patients report that the condition often

improves and worsens for no apparent reason. Although the exact cause of psoriasis

has not been established, many triggers associated with the onset and worsening of

the symptoms have been discovered in recent research. Established triggers include:

• stress

• skin injury (Koebner phenomenon)

• excessive alcohol consumption

• smoking

• changes of season or climate

• obesity

• streptococcal infection

• medications including Lithium, Antimalarials, Inderal, Quinidine and Indomethacin

6.1. PSORIASIS 69

A combination of genetic and environmental factors have been shown to be associ-

ated with the onset of psoriasis [Griffiths and Barker, 2007]. Almost 35% of psoriasis

patients report a family history of the disease, while monozygotic twins have been

shown to have a much higher concordance rate than dizygotic twins (65-73% versus

15-30%). Psoriasis has been found to be a typical complex disease in which both

genetic and environmental factors affect susceptability in family-based linkage and

epidemiological studies. The heritability of psoriasis, a measure of the proportion of

variability of the disease due to genetic factors, is estimated as 60-90% in Caucasians

[Elder et al., 1994]. Remarkably, this rate has been shown to be as high as 90-100%

in Danish twins[Brandrup and Green, 1981].

Traditional approaches to the identification of genetic risk factors, such as population-

based candidate gene association studies, have had moderate success. Associations

with markers in the major histocompatibility complex (MHC) region on chromosome

6 more were discovered over 35 years ago [Russell et al., 1972]. To date this locus

remains the major susceptibility locus for psoriasis, attributable directly to 35-50%

of Causasian genetic susceptibility for early-onset psoriasis. Strong associations have

been found between familial psoriasis and human leukocyte antigen (HLA) class I

genes, particularly HLA-Cw6. This has been shown to have a prevalence of up to

85% in early-onset patients compared with 15% in late-onset psoriasis and approxi-

mately 10% in the general population [Henseler and Christophers, 1985].

It is still unclear how these genes work together, with the main value of genetic

studies lying in the identification of molecular mechanisms and pathways for further

study. The majority of epidemiological research published to date has been concerned

with HLA-Cw6 exclusively as a genetic factor (with a multitude of environmental

factors), and the meta-analysis carried out in this thesis has concentrated exclusively

on this research. As further research is carried out into alternative candidate genes,

the methods developed here can incorporate these in the analysis with little extra

effort.

Contemporary genetic psoriasis research has already shown potential, with nine

locations(loci) of interest being found in single nucleotide polymorphism (SNP)-based

70 CHAPTER 6. PSORIASIS META-ANALYSIS

genome-wide association studies (GWASs). These are termed psoriasis susceptibil-

ity 1 through 9 (PSORS1 through PSORS9) [Nestle et al., 2009]. Within those loci

are genes, with HLA-Cw6 lying in the PSORS1 region. Many other of those genes

are on pathways that lead to inflammation. Technological advances leading to high-

throughput, accurate and simultaneous genotyping of hundreds of thousands of SNPs

has bought a new era of genetic studies in which the whole genome can be systemati-

cally screened in a hypothesis-free manner. This has the potential to uncover further

novel susceptibility markers in GWAS’s for psoriasis.

6.2 The data

The data set used in this analysis was gathered from 65 association studies in the

time period 1980 to 2008. Appendix A contains the references for the chosen papers.

Studies were included and excluded from this dataset in line with standard meta-

analysis procedures as outlined in [Stroup et al., 2000]. Linkage and family studies

were not considered at this stage of the analysis. We chose to include only studies

containing carrier frequencies, and replicated studies on same cohorts were excluded.

Each study was retrospective, and looked specifically at genotype information (HLA-

Cw6) plus a collection of other known risk factors, both demographic and clinical. In

addition to the case-control variable, the factors chosen for inclusion in this analysis

were the following:

• Gender (male or female)

• HLA-Cw6 (positive or negative)

• Ethnicity (Asian or Caucasian)

• Onset (early or late)

• Familial history (yes or no)

• Arthritis (yes or no)

6.2. THE DATA 71

• Type (vulgaris or guttate)

As an example to help explain some of the algorithmic details in Section 6.3, a

typical study such as Alenius et al. (2002) appears in the raw unprocessed data set

as Table 6.1. Please note that the control data contained in the table is not available

in the more granular early vs late familial history form.

n case n control male case male control female case female control1 88 84 37 512 713 17

ethnicity familial history onset arthritis type a1 case a1 control1 caucasian all yes 26 132 caucasian early yes 223 caucasian late yes 4

a2 case a2 control1 62 712 493 13

Table 6.1: Raw Data: Alenius et al. (2002)

In the printed format this table (shown here as three tables due to width restric-

tions) may look unwieldy, however it succinctly contains all of the data relevant to the

analysis. The blank spaces indicate missing information, with all three rows relating

to the same study. Multiple tables/slices of information are contained within this

single study and in fact the post processed data would appear as three small matrices

in Tables 6.2:

There is a considerable work required in order to join these three tables of infor-

mation in a coherent fashion, in order to produce the maximum likelihood estimate of

their combined table. Some details of this process are contained in the next section.

72 CHAPTER 6. PSORIASIS META-ANALYSIS

male femalecase 37 51

(a) Table 1

a1 a2control 13 71

(b) Table 2

onset a1 a2early 22 49late 4 13

(c) Table 3

Table 6.2: Processed Data: Alenius et al. (2002)

6.3 Methods

While the general methods involved in the EM algorithm have been previously ex-

plained (Chapter 2), some of the details of the data-specific problems have been

overlooked to date. In this section we will extrapolate upon the specific practical

mechanics involved in the E and first part of the M step of the overall algorithm. In

particular we shall concentrate on the five steps of how we go from tables such as

6.2 above for each study (with restricted subsets of the risk factors of interest) to an

overall estimate for the full-factor table. We have attempted to explain the concepts

herein without resorting to the use of mathematical notation, so as to appeal to a

larger audience. A more rigorous mathematical development is available in Chapter

2.

Level 1: This is the starting level of the analysis, often with multiple tables for each

study of differing sizes and shapes. Level 1 to Level 3 is the process by which the

within-study tables are combined and this is shown in a simplified picture in 6.1 To

achieve this we first found the smallest common table (SCT) for each study. This is

the minimum sized multidimensional array that contains all of the observed tables

from a particular study. Obviously each studies may have a different SCT.

Level 2: Each table has variables which are present, missing and conditional, when

6.3. METHODS 73

Figure 6.1: Multiple tables from a single study

compared to its SCT. The first step is to expand the tables from Level 1 across

their respective missing variables, in effect stretching the observed table across a

missing/marginal dimension. This step was not outlined in this context in Chapter

2, but is necessary when the slices are not disjoint, but rather multiple perspectives

on the same dimensions. Put more clearly, this step is necessary if we observe views

of the table which are overlapping; containing non-distinct variables.

Level 3: Therefore at Level 2 what remains are a collection of slices of the SCT, as the

expanded tables may contain conditional variables also. These are variables at which

we know some but not all of the information, e.g. a table with data from females

but not males. Modified SCT tables relating to each observed table are created by

filling out each slice into the current overall estimate of the SCT. In this way we can

mutually satisfy all of constraints of the constituent tables in a single study.

Level 4: Upon the completion of Level 3, the estimated SCT for each study is

74 CHAPTER 6. PSORIASIS META-ANALYSIS

Figure 6.2: Multiple studies to produce the full table

produced. To get an estimate for the overall full table, it is necessary to combine

the results of each of these studies. In a similar fashion to what was carried out to

produce the Level 2 tables, here we expand across the missing variables for each of the

studies (assume n of these in total). Figure 6.2 provides a low-dimensional pictorial

analogue of the process involved.

Level 5: Once again we have the issue with conditional variables as the tables

produced at Level 4 are in fact slices. In much the same manner as we did previously

we fill in each of these slices, producing a collection of n modified full tables. It is

these Level 5 tables which are used in the remainder of the M-step of the algorithm,

i.e. the retrospective adjustments and log-linear model fitting.

6.4. RESULTS 75

6.4 Results

6.4.1 Model fitting

The model-fitting process for these data sets must be approached with care. We have

already outlined that only interactions present in at least one constituent study may

appear in the overall M-step model, but we discovered many redundant model terms

even under these restrictions. For a full list of all potential terms, please see Appendix

B. This model had a G2 value of 10.19, with df = 96. The model-selection process

outlined in Section 2.10 allowed us to prune the final model appreciably, using the

G2 value and by testing the residual deviance.

Fitted Models G2 Residual Deviance df1 Saturated model 10.19 0 12 All linear terms 15.51 1253 2473 Case-control 25.88 14159 2544 Gender 52.37 44925 2545 Genetic 41.83 38573 2546 Ethnicity 41.83 38573 2547 Onset 52.39 38648 2548 Familial history 52.45 44923 2549 Arthritis 52.39 44926 254

10 Type 52.49 44886 25411 Intercept-only 52.44 44693 25512 Chosen Model 10.52 14 232

Table 6.3: G2 and residual deviances for 12 candidate loglinear models

Table 6.3 shows a set of 12 models, from the fully saturated model to the intercept-

only model, with a collection of models involving only single linear terms in between.

Contained within this table is also the associated G2, residual deviance and degrees

of freedom (df) for each of the model fits. As stated previously, an appropriate model

would have a low G2 and also low residual deviance, but with few fitted parameters

(a high df). What is immediately apparent is that the case-control variable is pivotal,

as it provides a large proportion of the reduction in the G2 value, when we compare

76 CHAPTER 6. PSORIASIS META-ANALYSIS

against the intercept-only model (∆G2 = 25.88−10.19 = 15.71 and ∆df = 254−1 =

253). Interestingly only moderate gains are made above this performance when we

include all of the linear terms (Model 2, with G2 = 15.51). While we will not go

through the full model-selection process, the most interesting thing perhaps was that

the linear terms for gender, familial history and onset were found to be redundant.

The chosen model was found to be (in R notation):

Frequency ∼ cc∗genetic∗ethnicity ∗arthritis+cc∗genetic∗ethnicity ∗ type. (6.1)

and as seen in Table 6.3, it has a G2 value close to the of the saturated model and

also reduced the residual deviance almost to zero. For those not familiar with the

modelling notation of R, equation 6.1 states the frequencies in the data sets are best

predicted by a model which includes the interactions cc∗genetic∗ethnicity∗arthritisand cc ∗ genetic ∗ ethnicity ∗ type, plus all of the corresponding lower order terms.

Thereafter we considered the convergence of the chosen model to ensure that the

algorithm was not getting stuck in local saddle points. We varied the initial distri-

bution within the interior parameter space, but the algorithm converged to the same

resulting distribution each time. Figure 6.3 shows the convergence of 20 randomly

chosen cells in the 8-way table, across 1000 iterations of the algorithm.

6.4.2 Testing homogeneity and finding influential studies

Tests for homogeneity are an integral part of any meta-analysis, in order to determine

whether the studies can reasonably be described as sharing a common effect size. We

carried out the test described in 2.8 and also the consider the jackknife influence

statistic for each of the studies. These are shown in Figure 6.4. It is immediately

apparent from this plots that Asumalahti et al. [2003], Chang et al. [2006] and Martin

et al. [2002] are studies of high leverage and warranted further attention. We revisited

these studies to confirm the data and investigate the reason for such high influence.

Asumalahti et al. [2003] influence derived from the fact that it was the larger of

only two studies which contained information on psoriasis gutate rather than psoriasis

vulgaris. Hence it had a large influence on disease prediction in this area of the table

6.4. RESULTS 77

Figure 6.3: Convergence for 20 elements of the Psoriasis estimated table

due to the sparsity in that region.

Asumalahti et al. [2003] would appear to be completely in order also, perhaps

with an odds ratio slightly higher than other contemporary studies, especially in the

early onset category, but still within confidence bounds. As this study had the lowest

jackknife influence score of the three studies in question, we decided to retain it in

the analysis.

Chang et al. [2006] was a study which contain only information about case-control

versus the genetic variable (HLA-Cw6) for arthritic psoriasis patients. The study size

78 CHAPTER 6. PSORIASIS META-ANALYSIS

Figure 6.4: Finding studies of high influence using the jackknife influence

was reasonably large (n = 650), and it had a value of 2.02 versus the fitted model’s

2.52, for the odds of psoriasis at the HLA-Cw6 positive versus negative. The reason

for the large influence of this study is it’s high degree of missingness, but the results

appears to be valid and hence it was included in the analysis.

6.4.3 Comparison against standard meta-analysis

We have already stated the advantages of the algorithms outlined in this thesis, versus

the more traditional meta-analysis methods. Namely, we have developed an omnibus

method which provides simultaneous predictions across a collection of risk factors,

rather than a single effect size as seen in more standard meta-analyses. The methods

proposed in this thesis also gain statistical power from combining so many studies in

this way. To compare the two methods and contrast their results, we shall look at

6.4. RESULTS 79

the most common effect size estimable from the 65 studies. There are five studies

[Economidou et al., 1985, Holm et al., 2003, Nair et al., 2008, O’Brien et al., 2001,

Sanchez et al., 2008], which contain information about case-control versus the genetic

variable for caucasians and are marginal to all of the other risk factors. Therefore,

the odds ratio for a1 (positive) versus a2 (negative) HLA-Cw6 may be calculated for

each of these studies, and both a fixed and a random effects model may be fitted.

The results of this analysis are found in Figure 6.5.

Figure 6.5: Combining 5 studies using fixed and random effect models

A Q-statistic of 11.35 is found with 4df (p-value = 0.023), suggesting a random

effects model is more appropriate in this instance. The estimated overall OR under

this model is 4.54, with a 95% CI of [3.19,6.45].

80 CHAPTER 6. PSORIASIS META-ANALYSIS

These results do not concur completely with those produced using our methods.

Using the likelihood-based approach, the estimated odds-ratio was 2.82, with a 95%

CI of [2.27,3.50]. Most obviously, this estimate is much lower than that produced by

the conventional meta-analysis. This may be accounted for by the fact that these

five studies were not the only contributary studies in finding the estimate under

our model, as many other studies contained information about this slice of the full

table. Upon further inspection, we found that many of the other studies contributed

lower OR estimates for this variable combination, with a mean OR of 1.65. Hence

the results found under our model were perhaps more reasonable, as the traditional

meta-analysis methods were not able to leverage these overlapping studies.

6.4.4 Disease prediction

The stated aim in building these models was to provide a method to estimate disease

probability for patients based on their individual risk characteristics. The major

advantage of the models we have fit, lies in the fact that they predict across the full

range of risk factors. In the psoriasis data set, we found four significant risk factors:

gene HLA-Cw6, ethnicity, arthritis and type. While the model itself includes many

higher order terms, to provide an overall picture for the relationship between each

of these risk factors and the disease probability, the marginal disease probabilities

for these variables are provided in Tables 6.4. The estimated odds ratios are also

included.

Factor Level1 Level2 OR1 Genetic (a1 vs a2) 0.033 0.015 2.212 Ethnicity (caucasian vs asian) 0.055 0.008 7.273 Arthritis (no vs yes) 0.018 0.022 0.824 Type (vulgaris vs gutate) 0.026 0.015 1.68

Table 6.4: Estimated marginal disease probabilities and odds-ratios

Please note that while there is a very high OR estimate for ethnicity, the CI is

also quite wide [1.25,8.62]. This is because only one moderately sized study contained

6.4. RESULTS 81

information for both asian and caucasian patients. This results in a imprecise estimate

across this margin.

The plots in Figure 6.6 show the estimated disease probability for four patients

with different risk characteristics, with associated 95% confidence intervals. In each

plot the dashed green line indicates the estimate disease rate of 0.02 for the entire

population.

Figure 6.6: Prediction intervals for four patients with different risk characteristics

Chapter 7

Alzheimer’s Disease Meta-Analysis

7.1 Alzheimer’s disease

Alzheimer’s disease (AD) is a progressive and fatal brain disorder first diagnosed by

a German physician named Alois Alzheimer in 1906. A degenerative and incurable

condition, it is the most common form of dementia, accounting for up to 70% of

cases. AD destroys brain cells, leading to memory and intellectual problems, severely

inhibiting the quality of life of those affected. For the majority of sufferers it is a late

onset disease (over 65 years), but an early onset version does also exist. It is estimated

that 26.6 million patients worldwide suffer from the disease, and this figure is expected

to quadruple before 2050. There is an accelerating worldwide effort under way to find

better ways to treat the disease, delay its onset, or prevent it from developing. No

succesful treatments have been found to delay onset or reduce disease risk, but it

remains an active research area for large pharmaceutical companies. For example,

in 2008 there were over 500 clinical studies investigating AD treatments. Although

the gestation period before diagnosis differs across patients, the mean life expectancy

is seven years. In fact fewer than 3% of sufferers survive more than fourteen years

following diagnosis.

While the cause and development of Alzheimer’s disease are not fully understood,

modern experts have linked the disease with plaques and tangles in the brain. This

link has been garnered predominantly based on autopsy evidence, and it is believed

82

7.1. ALZHEIMER’S DISEASE 83

that abnormally high growths of these structures disrupt, damage and kill nerve cells.

Some established and proposed risk factors for Alzheimer’s include:

• Age: the greatest known risk factor, with the likelihood of developing the disease

doubling every five years for those over 65.

• Family history: it has been found the risk of the disease doubles if a sibling has

previously suffered from the disease. Approximately 7% of disease incidence is

due to direct inheritance patterns associated with rare genes and regarded med-

ically as familial, i.e. passed on to 50% of the affected individual’s progeny. The

vast majority of these are attributable to mutations in one of three genes, amy-

loid precursor protein (APP) and presenilins 1 and 2 [Waring and Rosenberg,

2008].

• Aluminum: during the 1960s and 1970s it was suspected that regular exposure

to aluminum in household items such as pots, pans, beverage containers and

antiperspirants was linked to the onset of Alzheimer’s. Recent research has

failed to confirm this hypothesis, with few scientists still supporting this link.

• Head injury: serious cranial trauma has been shown to be linked directly with

the disease.

• General health: physically active and mentally stimulated individuals with a

healthy diet have been found to have both a lower rate of the disease and also

a later mean onset age for those who are diagnosed.

• Heart problems: the risk of developing the disease is increased by conditions

that damage the heart and blood vessels, such as diabetes, stroke, heart disease

and high cholesterol. Plaque and tangle growth has been shown to be greater

in AD sufferers who have heart issues.

• Genetic factors: the vast majority of cases of Alzheimer’s disease are sporadic

rather than familial, with genetic differences deigned only to be risk factors

here. The most established genetic link is the ε4 allele of the apolipoprotein

84 CHAPTER 7. ALZHEIMER’S DISEASE META-ANALYSIS

E (ApoE). This gene has been implicated in up to 50% of late onset sporadic

AD. Recent research has also linked angiotensin-converting enzyme (ACE) and

nitric oxide synthase (NOS3) to disease development, although this relationshop

has not yet been confirmed as conclusive. As noted for psoriasis genome-wide

association studies (GWASs) is unveiling avenues of discovery.

7.2 The data

The data set used in this analysis was gathered from 95 association studies in the

time period 1994 to 2006. Appendix B contains the references for the chosen papers.

Studies were included and excluded from this dataset in line with standard meta-

analysis procedures as outlined in [Stroup et al., 2000]. Linkage and family studies

were not considered at this stage of the analysis. We chose to include only studies

containing carrier frequencies, and replicated studies on same cohorts were excluded.

Each study was retrospective, and looked specifically at genotype information (ApoE,

ACE and NOS3) plus a collection of other known risk factors, both demographic and

clinical. In addition to the case-control variable, the factors chosen for inclusion in

this analysis were the following:

1. Gender: 2 levels (male and female)

2. ApoE: 6 levels, for each possible allele pair combination (ε2 ε2, ε2 ε3, ε2 ε4, ε3 ε3,

ε3 ε4 and ε4 ε4)

3. ACE: 3 levels, for each possible allele pair combination (ins ins, ins del and

del del)

4. NOS3: 3 levels, for each possible allele pair combination (glu glu, glu asp and

asp asp)

5. Ethnicity: 4 levels (caucasian, hispanic, asian and african)

6. Onset: 2 levels (early and late)

7. Familial history: 2 levels (familial or sporadic)

7.2. THE DATA 85

(a) ApoE (b) NOS3

Figure 7.1: ApoE and NOS3 structures

As an example of some of the data processing involved, plus in order to appreciate

the level of data fragmentation in the form of multiple slices, we have provided a look

at one of the constituent studies [Heinonen et al., 1995]. The raw data is provided in

Table 7.1, with the post-processed data provided in Table 7.2, with four slices/tables

of information emanating from this single study.

n case n control gene male case male con female case female con1 9 16 APOE 6 6 3 102 25 16 APOE 9 6 16 103 9 16 APOE 5 6 4 104 15 16 APOE 10 6 5 10

familial history onset ε2 ε2 case ε2 ε2 con ε2 ε3 case ε2 ε3 con ε2 ε4 case1 sporadic early 0 0 0 0 02 sporadic late 0 0 2 0 03 familial early 0 0 0 0 04 familial late 0 0 0 0 1

ε2 ε4 con ε3 ε3 case ε3 ε3 con ε3 ε4 case ε3 ε4 con ε4 ε4 case ε4 ε4 con1 0 2 13 4 2 3 12 0 5 13 14 2 4 13 0 5 13 1 2 3 14 0 7 13 4 2 3 1

Table 7.1: Raw Data: Lehtovirta et al. (1996)

86 CHAPTER 7. ALZHEIMER’S DISEASE META-ANALYSIS

male femalecontrol 6 10

(a) Table 1: Male and femalecontrols

early.sporadic late.sporadic early.familial late.familialmale 6 9 5 10

female 3 16 4 5

(b) Table 2: Gender vs onset vs familial history cases

controlε2 ε2 0ε2 ε3 0ε2 ε4 0ε3 ε3 13ε3 ε4 2ε4 ε4 1

(c) Table 3: ApoEcontrols

early.sporadic late.sporadic early.familial late.familialε2 ε2 0 0 0 0ε2 ε3 0 2 0 0ε2 ε4 0 0 0 1ε3 ε3 2 5 5 7ε3 ε4 4 14 1 4ε4 ε4 3 4 3 3

(d) Table 4: ApoE vs onset vs familial history cases

Table 7.2: Processed Data: Lehtovirta et al. (1996)

The Alzheimer disease data set contained many random and structural zeros, as

seen in the tables above. It is therefore important to take care in the implementation

of the algorithm, using the methods for dealing with zeros outlined in Agresti [2002];

introducing a constraint in the initial full table estimate.

It should also be noted that each of these tables contains information on either

the cases or controls. The control data was not stratified across the onset and familial

history variables. The complexity of data structure observed in this study is typical

7.3. RESULTS 87

of the Alzheimer’s disease data base.

7.3 Results

As previously seen in the psoriasis analysis, the model-fitting process for these data

sets must be carried out with care. For a full list of all potential terms contained in at

least one study, please see Appendix D. In the AD analysis none of the linear terms

were excluded and the chosen model is included in D also. We once again checked for

aberrant/influential studies, to verify the validity of the constituent studies. Three

suspicious studies were highlighted, namely Molero et al. [2001], Romas et al. [2002],

Wakutani et al. [2002] and these were investigated further.

Figure 7.2: Finding aberrant studies in the Alzheimer’s disease data set

The high influence of Molero et al. [2001] was due to its large sample size (n=1785),

88 CHAPTER 7. ALZHEIMER’S DISEASE META-ANALYSIS

and hence we have chosen to retain this study in the analysis.

Romas et al. [2002] was yet another study of a Latin population, but with a

particularly large odds ratio for the ε3 ε3 versus ε2 ε2 of 2.97. There does not seem

to anything peculiar in the research itself, but this effect size is outside the range of

the estimate upon its exclusion. Therefore this study has been excluded from the

remainder of the analysis.

As the sole japanese study in this analysis, Wakutani et al. [2002] differs in its

observed rates than those of the other Asian studies it is pooled with, which otherwise

concentrate on predominantly Chinese populations. Again this does not appear to

be an homogeneous study and is excluded until such time as other Japanese studies

are included in the data base.

It is difficult to provide plots which give an appreciable perspective of such a high

dimensional model. In Figure 7.3 we show odds ratios relating to the four margins of

the fitted model, for gender and the three genetic variables. Of course it is important

to note that each of these plots looks at the variable effect summed across the other

variables levels, while in reality the model contains many interactions hidden from

these figures. However, even after considering this fact, the results shown in these

plots are exceptionally interesting. Each plot contained the OR estimate against the

chosen baseline, with a 95% confidence interval provided.

Firstly, the higher risk of AD is confirmed for females, and the well established

ε4 (a3) risk in ApoE was immediately obvious in second plot also. Perhaps the most

interesting conclusions are highlighted in the remaining two plots. Here we see that

ACE and NOS3 both have significant odds-ratio, indicating that they do indeed have

a link with the onset of Alzheimer’s disease. This relationship was suspected but not

confirmed in previous literature.

Two prediction plots, with their associated confidence intervals, are provided in

Figure 7.4 for patients with differing risk characteristics. Plots such as this will inform

patients and doctors in IVF clinics, and guide decisions on further genetic screening.

7.3. RESULTS 89

(a) Gender Marginal (b) ApoE Marginal

(c) ACE Marginal (d) NOS3 Marginal

Figure 7.3: Estimated marginal distributions gender and the three genetic risk factors

90 CHAPTER 7. ALZHEIMER’S DISEASE META-ANALYSIS

(a) Patient 1: Late Onset

(b) Patient 2: Early Onset

Figure 7.4: Prediction intervals for two patients with specific risk characteristic loci

Chapter 8

Conclusions

The main contributions of this thesis are as follows:

• A likelihood-based model was developed for the synthesis of many partially

classified contingency tables, using the EM algorithm and loglinear models.

This algorithm searches each constituent table and includes only interactions

observed in at least one table in the overall model. We have shown this method

to be consistent and accurate.

• We have also proposed a second, Bayesian model. This algorithm uses the anal-

ogous Data Augmentation algorithm in conjunction with Bayesian IPF, and

produces estimates close to those of the likelihood-based model. It does how-

ever have the beneficial property of providing an estimate for the full posterior

distribution, rather than just a point estimate for the mean as we saw in the

earlier model.

• The missing data patterns in this thesis were complex and varied. We provided

solutions for dealing with missingness due to marginal and conditional variables,

and combinations thereof, for both the likelihood-based setting and the Bayesian

model.

• Each study in the meta-analysis was case-control. We extended the previous

work on adjustments for retrospective sampling, finding the suitable correction

91

92 CHAPTER 8. CONCLUSIONS

under the logistic model for multiple studies. The loglinear case was not ex-

plored in previous research, but here we provided the adjustments necessary for

the single and multiple study scenario under this model.

• A multitude of variance estimates were investigated, including parametric meth-

ods based on the sandwich estimate, the bootstrap, the jackknife, SEM and the

posterior variance. The jackknife was found to out-perform the parametric

methods, while also providing supplementary benefits including natural cross-

validation samples for model-selection and in finding influential studies.

• Two meta-analysis were presented in this dissertation, using data sets collected

for psoriasis and Alzheimer’s disease. Predictive models were found using the

techniques developed in this thesis on a wide range of risk factors. These are

the largest and most accurate predictive models found to date for these high

profile diseases, confirming some hitherto unproven hypotheses regarding risk

factors using the increased power of the aggregated samples.

Appendix A

Studies in the psoriasis data base

The following studies were contained in the psoriasis data base: Al-Heresh et al.

[2002], Alenius et al. [2002], Allen et al. [2005], Armstrong et al. [1983], Asahina

et al. [1991], Asumalahti et al. [2000, 2003], Atasoy et al. [2006], Brenner et al. [1978],

Chang et al. [2003a,b, 2006], Choi et al. [2000], Dobosz et al. [2005], Duffin and

Krueger [2009], Economidou et al. [1985], Fan et al. [2007], Fojtikova et al. [2009],

Fry et al. [2006], Gladman et al. [1999], Gonzalez et al. [2001, 2000, 1999], Gudjonsson

et al. [2003], Helms et al. [2005], Ho et al. [2008], Hohler et al. [1996], Holm et al.

[2003, 2005a,b], Ikaheimo et al. [1996], Jobim et al. [2008], Kastelan et al. [2000], Kim

et al. [2000b], Kundakci et al. [2002], Liao et al. [2008], Lopez-Larrea et al. [1990],

Luszczek et al. [2003], Mallon et al. [2000, 1997, 1998], Martin et al. [2002], Martinez-

Borra et al. [2003], Murray et al. [1980], Nair et al. [2008], Nakagawa et al. [1991],

O’Brien et al. [2001], Orru et al. [2002], Ozawa et al. [1988], Pyo et al. [2003], Queiro

et al. [2008, 2006, 2003], Queiro-Silva et al. [2004], Rahman et al. [2003], Rani et al.

[1998], Roitberg-Tambur et al. [1994], Romphruk et al. [2003], Sanchez et al. [2004,

2008], Schmitt-Egenolf et al. [1996], Szczerkowska-Dobosz et al. [2004], Vejbaesya

et al. [1998], Williams et al. [2005], Wisniewski et al. [2003]

93

Appendix B

Interactions present in psoriasis

studies

A full list of the interactions present in at least one psoriasis model is given by the

following:

cc+ gender+ genetic+ ethnicity+ onset+ familialhistory+ arthritis+ type+

gender ∗ cc ∗ ethnicity ∗ arthritis+ cc ∗ genetic ∗ ethnicity ∗ arthritis+ genetic ∗ cc ∗ethnicity∗arthritis+onset∗genetic∗cc∗ethnicity∗arthritis+gender∗cc∗ethnicity∗onset∗type+cc∗genetic∗ethnicity∗onset∗type+genetic∗cc∗ethnicity+arthritis∗genetic∗ cc∗ethnicity+gender ∗ cc∗ethnicity ∗ type+ cc∗genetic∗ethnicity ∗ type+

onset∗genetic∗cc∗ethnicity+cc∗genetic∗type+cc∗gender∗ethnicity∗type+genetic∗cc∗ethnicity ∗ type+onset∗genetic∗cc∗ethnicity ∗ type+cc∗genetic∗onset∗ type+

gender ∗ cc ∗ ethnicity+ cc ∗ genetic ∗ ethnicity+ cc ∗ genetic+ gender ∗ genetic ∗ cc ∗ethnicity+cc∗genetic∗ethnicity∗onset∗arthritis+type∗gender∗cc∗ethnicity+type∗genetic∗cc∗ethnicity+gender ∗cc∗ethnicity ∗onset∗arthritis∗ type+cc∗genetic∗ethnicity ∗ onset ∗ arthritis ∗ type+ arthritis ∗ gender ∗ cc ∗ ethnicity+ cc ∗ gender ∗ethnicity∗onset∗type+genetic∗cc∗ethnicity∗onset+gender∗genetic∗cc∗ethnicity∗onset+ cc ∗ gender ∗ type+ genetic ∗ cc ∗ type+ onset ∗ genetic ∗ cc ∗ type+ genetic ∗familialhistory∗onset∗cc∗ethnicity∗type+genetic∗onset∗gender∗cc∗ethnicity∗type+ cc ∗ genetic ∗ arthritis+ genetic ∗ cc ∗ onset+ arthritis ∗ genetic ∗ cc ∗ onset

94

Appendix C

Studies in the Alzheimer’s data

base

The following studies were contained in the Alzheimer’s disease data base:

Adroer et al. [1995],Alvarez et al. [1999],Alvarez-Alvarez et al. [2003],Arboleda

et al. [2001],Bang et al. [2003],Buss et al. [2002],Cacabelos et al. [2003],Camelo et al.

[2004],Carrieri et al. [2001],Chen et al. [1999a],Chen et al. [1999b],Cheng et al. [2002],Corder

et al. [1994],Crawford et al. [2000],Cui et al. [2000],Farrer et al. [2000],Graff-Radford

et al. [2002],Guidi et al. [2005],Heinonen et al. [1995],Higuchi et al. [2000],Hong et al.

[1996],Hu et al. [1999],Huang et al. [2002],Isbir et al. [2000],Juhasz et al. [2005],Kim

et al. [2000a],Kim et al. [2001],Kolsch et al. [2005],Kukull et al. [1996],Kunugi et al.

[2000],Lambert et al. [1998],Lannfelt et al. [1994],Lehtovirta et al. [1996],Lendon et al.

[2002],Liu et al. [1999],Lopez et al. [1998],Ma et al. [2005],Maestre et al. [1995],Molero

et al. [2001],Monastero et al. [2002],Monastero et al. [2003],Mui et al. [1996],Myl-

lykangas et al. [2000],Nakayama and Kuzuhara [1999],Nalbantoglu et al. [1994],Narain

et al. [2000],Nunomura et al. [1996],Osuntokun et al. [1995],Panza et al. [2000],Panza

et al. [2003],Perry et al. [2001],Poirier et al. [1993],Prince et al. [2001],Quiroga et al.

[1999],Raygani et al. [2005],Richard et al. [2001],Romas et al. [2002],Roses [1997],Sa-

hota et al. [1997],Sanchez-Guerra et al. [2001],Scacchi et al. [1995],Scott et al. [1997],Seripa

et al. [2003],Seripa et al. [2004],Singleton et al. [2001],Sleegers et al. [2005],Slooter

et al. [1998],Sorbi et al. [1994],Souza et al. [2003],Styczynska et al. [2003],Sulkava

95

96 APPENDIX C. STUDIES IN THE ALZHEIMER’S DATA BASE

et al. [1996],Sunderland et al. [2004],Talbot et al. [1994],Tang et al. [1998],Tapi-

ola et al. [1998],Tedde et al. [2002],Tilley et al. [1999],Town et al. [1998],Tsuang

et al. [2005],van Duijn et al. [1995],Vuletic et al. [2005],Wakutani et al. [2002],Wang

et al. [2000],Wang et al. [2006],Wiebusch et al. [1999],Yang et al. [2000],Yang et al.

[2003],Zambenedetti et al. [2003],Zhang et al. [2003] and Zuliani et al. [2001]

Appendix D

Interactions present in Alzheimer’s

studies

A full list of the interactions present in at least one Alzheimer’s disease study is given

by the following:

cc + gender + APOE + ACE + NOS3 + ethnicity + onset + familialhistory +

cc ∗ gender ∗ ethnicity+ cc ∗APOE ∗ ethnicity+ gender ∗ cc ∗ ethnicity ∗ onset+ cc ∗ACE ∗ethnicity∗onset+cc∗gender∗ethnicity+cc∗APOE ∗ethnicity+cc∗gender∗ethnicity ∗onset+cc∗ACE ∗ethnicity ∗onset+cc∗APOE+cc∗APOE ∗ethnicity ∗familialhistory+gender∗cc∗onset+cc∗ACE∗onset+cc∗ACE∗onset+cc∗APOE∗ethnicity ∗onset+cc∗gender∗ethnicity ∗onset+cc∗APOE ∗ethnicity ∗onset+cc∗gender ∗ ethnicity+ cc∗ACE ∗ ethnicity+ cc∗APOE ∗ ethnicity ∗familialhistory ∗onset+ cc∗ACE ∗ethnicity ∗familialhistory+ cc∗APOE ∗ethnicity+ cc∗gender ∗ethnicity+cc∗ACE∗ethnicity+cc∗gender∗ethnicity∗onset+cc∗ACE∗ethnicity∗onset+cc∗gender∗ethnicity∗onset+cc∗ACE∗ethnicity∗familialhistory∗onset+cc∗APOE∗ethnicity+gender∗cc∗ethnicity∗familialhistory+cc∗ACE∗ethnicity∗familialhistory+ cc ∗ gender ∗ ethnicity ∗ familialhistory+ cc ∗NOS3 ∗ ethnicity ∗familialhistory+ cc∗ACE ∗ ethnicity+ cc∗NOS3∗ ethnicity ∗onset+ cc∗gender ∗ethnicity ∗ familialhistory ∗ onset+ cc ∗APOE ∗ ethnicity ∗ familialhistory ∗ onset

97

Bibliography

R Adroer, P Santacruz, R Blesa, S Lopez-Pousa, C Ascaso, and R Oliva. Apolipopro-

tein E4 allele frequency in Spanish Alzheimer and control cases. Neurosci Lett, 189

(3):182–6, 1995.

A. Agresti. Categorical Data Analysis. Wiley, New York, 2nd edition, 2002.

A M Al-Heresh, J Proctor, S M Jones, J Dixey, B Cox, K Welsh, and N McHugh. Tu-

mour necrosis factor-alpha polymorphism and the HLA-Cw*0602 allele in psoriatic

arthritis. Rheumatology (Oxford), 41(5):525–30, 2002.

G-M Alenius, E Jidell, L Nordmark, and S Rantapaa Dahlqvist. Disease manifesta-

tions and HLA antigens in psoriatic arthritis in northern Sweden. Clin Rheumatol,

21(5):357–62, 2002.

Michael Hugh Allen, Hahreen Ameen, Colin Veal, Julie Evans, V S Ramrakha-Jones,

A M Marsland, A David Burden, C E M Griffiths, Richard C Trembath, and

Jonathan N W N Barker. The major psoriasis susceptibility locus PSORS1 is not

a risk factor for late-onset psoriasis. J Invest Dermatol, 124(1):103–6, 2005.

R Alvarez, V Alvarez, C H Lahoz, C Martinez, J Pena, J M Sanchez, L M Guisasola,

J Salas-Puig, G Moris, J A Vidal, R Ribacoba, B B Menes, D Uria, and E Coto.

Angiotensin converting enzyme and endothelial nitric oxide synthase DNA poly-

morphisms and late onset Alzheimer’s disease. J Neurol Neurosurg Psychiatry, 67

(6):733–6, 1999.

Maite Alvarez-Alvarez, Luis Galdos, Manuel Fernandez-Martinez, Fernando Gomez-

Busto, Victoria Garcia-Centeno, Caridad Arias-Arias, Carmen Sanchez-Salazar,

98

BIBLIOGRAPHY 99

Ana Belen Rodriguez-Martinez, Juan Jose Zarranz, and Marian M de Pancorbo.

5-Hydroxytryptamine 6 receptor (5-HT(6)) receptor and apolipoprotein E (ApoE)

polymorphisms in patients with Alzheimer’s disease in the Basque Country. Neu-

rosci Lett, 339(1):85–7, 2003.

J. A. Anderson and S. C. Richardson. Logistic discrimination and bias correction in

maximum likelihood estimation. Technometrics, 21(1):71–78, 1979. ISSN 00401706.

URL http://www.jstor.org/stable/1268582.

G H Arboleda, J J Yunis, R Pardo, C M Gomez, D Hedmont, G Arango, and H Ar-

boleda. Apolipoprotein E genotyping in a sample of Colombian patients with

Alzheimer’s disease. Neurosci Lett, 305(2):135–8, 2001.

R D Armstrong, G S Panayi, and K I Welsh. Histocompatibility antigens in psoriasis,

psoriatic arthropathy, and ankylosing spondylitis. Ann Rheum Dis, 42(2):142–6,

1983.

A Asahina, S Akazaki, H Nakagawa, S Kuwata, K Tokunaga, Y Ishibashi, and T Juji.

Specific nucleotide sequence of HLA-C is strongly associated with psoriasis vulgaris.

J Invest Dermatol, 97(2):254–8, 1991.

K Asumalahti, T Laitinen, R Itkonen-Vatjus, M L Lokki, S Suomela, E Snellman,

U Saarialho-Kere, and J Kere. A candidate gene for psoriasis near HLA-C, HCR

(Pg8), is highly polymorphic with a disease-associated susceptibility allele. Hum

Mol Genet, 9(10):1533–42, 2000.

Kati Asumalahti, Mahreen Ameen, Sari Suomela, Eva Hagforsen, Gerd Michaelsson,

Julie Evans, Margo Munro, Colin Veal, Michael Allen, Joyce Leman, A David

Burden, Brian Kirby, Maureen Connolly, Christopher E M Griffiths, Richard C

Trembath, Juha Kere, Ulpu Saarialho-Kere, and Jonathan N W N Barker. Genetic

analysis of PSORS1 distinguishes guttate psoriasis and palmoplantar pustulosis. J

Invest Dermatol, 120(4):627–32, 2003.

Mustafa Atasoy, Ibrahim Pirim, Omer F Bayrak, Sevki Ozdemir, Mevlit Ikbal, Teo-

man Erdem, and Akin Aktas. Association of HLA class I and class II alleles with

100 BIBLIOGRAPHY

psoriasis vulgaris in Turkish population. Influence of type I and II psoriasis. Saudi

Med J, 27(3):373–6, 2006.

Oh Young Bang, Yong Tae Kwak, In Soo Joo, and Kyoon Huh. Important link

between dementia subtype and apolipoprotein E: a meta-analysis. Yonsei Med J,

44(3):401–13, 2003.

F Brandrup and A Green. The prevalence of psoriasis in Denmark. Acta Derm

Venereol, 61(4):344–6, 1981.

W Brenner, F Gschnait, and W R Mayr. HLA B13, B17, B37 and Cw6 in psoriasis

vulgaris: association with the age of onset. Arch Dermatol Res, 262(3):337–9, 1978.

Svenja Buss, Tomas Muller-Thomsen, Cristoph Hock, Antonella Alberici, Giuliano

Binetti, Roger M Nitsch, Andreas Gal, and Ulrich Finckh. No association between

DCP1 genotype and late-onset Alzheimer disease. Am J Med Genet, 114(4):440–5,

2002.

Ramon Cacabelos, Lucia Fernandez-Novoa, Valter Lombardi, Lola Corzo, Victor

Pichel, and Yasuhiko Kubota. Cerebrovascular risk factors in Alzheimer’s dis-

ease: brain hemodynamics and pharmacogenomic implications. Neurol Res, 25(6):

567–80, 2003.

Dalila Camelo, Gonzalo Arboleda, Juan J Yunis, Rodrigo Pardo, Gabriel Arango, Eu-

genia Solano, Luis Lopez, Daniel Hedmont, and Humberto Arboleda. Angiotensin-

converting enzyme and alpha-2-macroglobulin gene polymorphisms are not associ-

ated with Alzheimer’s disease in Colombian patients. J Neurol Sci, 218(1-2):47–51,

2004.

G Carrieri, M Bonafe, M De Luca, G Rose, O Varcasia, A Bruni, R Maletta,

B Nacmias, S Sorbi, F Corsonello, E Feraco, K F Andreev, A I Yashin, C Franceschi,

and G De Benedictis. Mitochondrial DNA haplogroups and APOE4 allele are non-

independent variables in sporadic Alzheimer’s disease. Hum Genet, 108(3):194–8,

2001.

BIBLIOGRAPHY 101

Y T Chang, S F Tsai, D D Lee, Y M Shiao, C Y Huang, H N Liu, W J Wang,

and C K Wong. A study of candidate genes for psoriasis near HLA-C in Chinese

patients with psoriasis. Br J Dermatol, 148(3):418–23, 2003a.

Y T Chang, S F Tsai, M W Lin, H N Liu, D D Lee, Y M Shiao, P J Chin, and W J

Wang. SPR1 gene near HLA-C is unlikely to be a psoriasis susceptibility gene. Exp

Dermatol, 12(3):307–14, 2003b.

Yun-Ting Chang, Chan-Te Chou, Yu-Ming Shiao, Ming-Wei Lin, Chia-Wen Yu, Chih-

Chiang Chen, Cheng-Hung Huang, Ding-Dar Lee, Han-Nan Liu, Wen-Jen Wang,

and Shih-Feng Tsai. The killer cell immunoglobulin-like receptor genes do not confer

susceptibility to psoriasis vulgaris independently in Chinese. J Invest Dermatol, 126

(10):2335–8, 2006.

L Chen, L Baum, H K Ng, L Y Chan, and C P Pang. Apolipoprotein E genotype and

its pathological correlation in Chinese Alzheimer’s disease with late onset. Hum

Pathol, 30(10):1172–7, 1999a.

L Chen, L Baum, H K Ng, L Y Chan, and C P Pang. Apolipoprotein E genotype and

its pathological correlation in Chinese Alzheimer’s disease with late onset. Hum

Pathol, 30(10):1172–7, 1999b.

Chih-Ya Cheng, Chen-Jee Hong, Hsiu-Chih Liu, Tsung-Yun Liu, and Shih-Jen Tsai.

Study of the association between Alzheimer’s disease and angiotensin-converting

enzyme gene polymorphism using DNA from lymphocytes. Eur Neurol, 47(1):26–9,

2002.

H B Choi, H Han, J I Youn, T Y Kim, and T G Kim. MICA 5.1 allele is a susceptibility

marker for psoriasis in the Korean population. Tissue Antigens, 56(6):548–50, 2000.

H. Cooper and L.V. Hedges. The Handbook of Research Synthesis. Russell Sage

Foundation, New York, 1994.

E H Corder, A M Saunders, N J Risch, W J Strittmatter, D E Schmechel, P C Jr

Gaskell, J B Rimmler, P A Locke, P M Conneally, and K E Schmader. Protective

102 BIBLIOGRAPHY

effect of apolipoprotein E type 2 allele for late onset Alzheimer disease. Nat Genet,

7(2):180–4, 1994.

F Crawford, L Abdullah, J Schinka, Z Suo, M Gold, R Duara, and M Mullan. Gender-

specific association of the angiotensin converting enzyme gene with Alzheimer’s

disease. Neurosci Lett, 280(3):215–9, 2000.

T Cui, X Zhou, W Jin, F Zheng, and X Cao. Gene polymorphism in apolipoprotein

E and presenilin-1 in patients with late-onset Alzheimer’s disease. Chin Med J

(Engl), 113(4):340–4, 2000.

A.P. Dempster, N.M. Laird, and D.B. Rubin. Maximum likelihood from incomplete

data via the em algorithm. Journal of the Royal Statistical Society Series B, B(39):

1–38, 1977.

A Szczerkowska Dobosz, K Rebala, Z Szczerkowska, and B Nedoszytko. HLA-C

locus alleles distribution in patients from northern Poland with psoriatic arthritis–

preliminary report. Int J Immunogenet, 32(6):389–91, 2005.

Kristina Callis Duffin and Gerald G Krueger. Genetic variations in cytokines and

cytokine receptors associated with psoriasis found by genome-wide association. J

Invest Dermatol, 129(4):827–33, 2009.

J Economidou, C Papasteriades, M Varla-Leftherioti, A Vareltzidis, and J Stratigos.

Human lymphocyte antigens A, B, and C in Greek patients with psoriasis: relation

to age and clinical expression of the disease. J Am Acad Dermatol, 13(4):578–82,

1985.

Bradley Efron and Robert J. Tibshirani. An Introduction to the Bootstrap. Chapman

and Hall, London, 1993.

J T Elder, R P Nair, S W Guo, T Henseler, E Christophers, and J J Voorhees. The

genetics of psoriasis. Arch Dermatol, 130(2):216–24, 1994.

BIBLIOGRAPHY 103

Xing Fan, Sen Yang, Liang Dan Sun, Yan Hua Liang, Min Gao, Kai Yue Zhang,

Wei Huang, and XueJun Zhang. Comparison of clinical features of HLA-Cw*0602-

positive and -negative psoriasis patients in a Han Chinese population. Acta Derm

Venereol, 87(4):335–40, 2007.

L A Farrer, T Sherbatich, S A Keryanov, G I Korovaitseva, E A Rogaeva, S Petruk,

S Premkumar, Y Moliaka, Y Q Song, Y Pei, C Sato, N D Selezneva, S Voskre-

senskaya, V Golimbet, S Sorbi, R Duara, S Gavrilova, P H St George-Hyslop, and

E I Rogaev. Association between angiotensin-converting enzyme and Alzheimer

disease. Arch Neurol, 57(2):210–4, 2000.

Marketa Fojtikova, Jiri Stolfa, Peter Novota, Pavlina Cejkova, Ctibor Dostal, and

Marie Cerna. HLA-Cw*06 class I region rather than MICA is associated with

psoriatic arthritis in Czech population. Rheumatol Int, 29(11):1293–9, 2009.

L Fry, A V Powles, S Corcoran, S Rogers, J Ward, and D J Unsworth. HLA Cw*06

is not essential for streptococcal-induced psoriasis. Br J Dermatol, 154(5):850–3,

2006.

Camil Fuchs. Maximum likelihood estimation and model selection in contingency

tables with missing data. Journal of the American Statistical Association, 77(378):

270–278, 1982. ISSN 01621459. URL http://www.jstor.org/stable/2287230.

A. Gelman, D.B Rubin, J. Carlin, and H. Stern. Bayesian Data Analysis. Chapman

and Hall, London, 1995.

Stuart German and Donald German. Stochastic relaxation, gibbs distributions, and

the bayesian restoration of images. IEEE Transactions on Pattern Analysis and

Machine Intelligence, PAMI-6(6):721–741, Nov. 1984.

D D Gladman, C Cheung, C M Ng, and J A Wade. HLA-C locus alleles in patients

with psoriatic arthritis (PsA). Hum Immunol, 60(3):259–61, 1999.

S Gonzalez, J Martinez-Borra, J C Torre-Alonso, S Gonzalez-Roces, J Sanchez del

Rio, A Rodriguez Perez, C Brautbar, and C Lopez-Larrea. The MICA-A9 triplet

104 BIBLIOGRAPHY

repeat polymorphism in the transmembrane region confers additional susceptibility

to the development of psoriatic arthritis and is independent of the association of

Cw*0602 in psoriasis. Arthritis Rheum, 42(5):1010–6, 1999.

S Gonzalez, J Martinez-Borra, J S Del Rio, J Santos-Juanes, A Lopez-Vazquez,

M Blanco-Gelaz, and C Lopez-Larrea. The OTF3 gene polymorphism confers sus-

ceptibility to psoriasis independent of the association of HLA-Cw*0602. J Invest

Dermatol, 115(5):824–8, 2000.

S Gonzalez, C Brautbar, J Martinez-Borra, A Lopez-Vazquez, R Segal, M A Blanco-

Gelaz, C D Enk, C Safriman, and C Lopez-Larrea. Polymorphism in MICA rather

than HLA-B/C genes is associated with psoriatic arthritis in the Jewish population.

Hum Immunol, 62(6):632–8, 2001.

Neill R Graff-Radford, Robert C Green, Rodney C P Go, Michael L Hutton, Timi

Edeki, David Bachman, Jennifer L Adamson, Patrick Griffith, Floyd B Willis, Mary

Williams, Yvonne Hipps, Jonathan L Haines, L Adrienne Cupples, and Lindsay A

Farrer. Association between apolipoprotein E genotype and Alzheimer disease in

African American subjects. Arch Neurol, 59(4):594–600, 2002.

Christopher E M Griffiths and Jonathan N W N Barker. Pathogenesis and clinical

features of psoriasis. Lancet, 370(9583):263–71, 2007.

J E Gudjonsson, A Karason, A Antonsdottir, E H Runarsdottir, V B Hauksson,

R Upmanyu, J Gulcher, K Stefansson, and H Valdimarsson. Psoriasis patients

who are homozygous for the HLA-Cw*0602 allele have a 2.5-fold increased risk of

developing psoriasis compared with Cw6 heterozygotes. Br J Dermatol, 148(2):

233–5, 2003.

Ilaria Guidi, Daniela Galimberti, Eliana Venturelli, Carlo Lovati, Roberto Del Bo,

Chiara Fenoglio, Alberto Gatti, Roberto Dominici, Sara Galbiati, Roberta Vir-

gilio, Simone Pomati, Giacomo P Comi, Claudio Mariani, Gianluigi Forloni, Nereo

Bresolin, and Elio Scarpini. Influence of the Glu298Asp polymorphism of NOS3

BIBLIOGRAPHY 105

on age at onset and homocysteine levels in AD patients. Neurobiol Aging, 26(6):

789–94, 2005.

T.J. Hastie, R.J. Tibshirani, and J.H. Friedman. The Elements of Statistical Learning.

Springer, New York, 2001.

Walter W. Hauck. The large sample variance of the mantel-haenszel estimator of

a common odds ratio. Biometrics, 35(4):817–819, 1979. ISSN 0006341X. URL

http://www.jstor.org/stable/2530114.

Larry V. Hedges and Ingram Olkin. Statistical Methods for Meta-Analysis. Academic

Press,San Diego, 1985.

O Heinonen, M Lehtovirta, H Soininen, S Helisalmi, A Mannermaa, H Sorvari, O Ko-

sunen, L Paljarvi, M Ryynanen, and P J Sr Riekkinen. Alzheimer pathology of

patients carrying apolipoprotein E epsilon 4 allele. Neurobiol Aging, 16(4):505–13,

1995.

Cynthia Helms, Nancy L Saccone, Li Cao, Jil A Wright Daw, Kai Cao, Tony M Hsu,

Patricia Taillon-Miller, Shenghui Duan, Derek Gordon, Brandon Pierce, Jurg Ott,

John Rice, Marcelo A Fernandez-Vina, Pui-Yan Kwok, Alan Menter, and Anne M

Bowcock. Localization of PSORS1 to a haplotype block harboring HLA-C and

distinct from corneodesmosin and HCR. Hum Genet, 118(3-4):466–76, 2005.

T Henseler and E Christophers. Psoriasis of early and late onset: characterization of

two types of psoriasis vulgaris. J Am Acad Dermatol, 13(3):450–6, 1985.

S Higuchi, S Ohta, S Matsushita, T Matsui, T Yuzuriha, K Urakami, and H Arai.

NOS3 polymorphism not associated with Alzheimer’s disease in Japanese. Ann

Neurol, 48(4):685, 2000.

P Y P C Ho, A Barton, J Worthington, D Plant, C E M Griffiths, H S Young,

P Bradburn, W Thomson, A J Silman, and I N Bruce. Investigating the role

of the HLA-Cw*06 and HLA-DRB1 genes in susceptibility to psoriatic arthritis:

106 BIBLIOGRAPHY

comparison with psoriasis and undifferentiated inflammatory arthritis. Ann Rheum

Dis, 67(5):677–82, 2008.

T Hohler, A Weinmann, P M Schneider, C Rittner, R E Schopf, J Knop,

P Hasenclever, K H Meyer zum Buschenfelde, and E Marker-Hermann. TAP-

polymorphisms in juvenile onset psoriasis and psoriatic arthritis. Hum Immunol,

51(1):49–54, 1996.

Sofia J Holm, Lina M Carlen, Lotus Mallbris, Mona Stahle-Backdahl, and Kevin P

O’Brien. Polymorphisms in the SEEK1 and SPR1 genes on 6p21.3 associate with

psoriasis in the Swedish population. Exp Dermatol, 12(4):435–44, 2003.

Sofia J Holm, Kazuko Sakuraba, Lotus Mallbris, Katarina Wolk, Mona Stahle, and

Fabio O Sanchez. Distinct HLA-C/KIR genotype profile associates with guttate

psoriasis. J Invest Dermatol, 125(4):721–30, 2005a.

Sofia J Holm, Fabio Sanchez, Lina M Carlen, Lotus Mallbris, Mona Stahle, and

Kevin P O’Brien. HLA-Cw*0602 associates more strongly to psoriasis in the

Swedish population than variants of the novel 6p21.3 gene PSORS1C3. Acta Derm

Venereol, 85(1):2–8, 2005b.

C J Hong, T Y Liu, H C Liu, S J Wang, J L Fuh, C W Chi, K Y Lee, and C B

Sim. Epsilon 4 allele of apolipoprotein E increases risk of Alzheimer’s disease in a

Chinese population. Neurology, 46(6):1749–51, 1996.

J Hu, F Miyatake, Y Aizu, H Nakagawa, S Nakamura, A Tamaoka, R Takahash,

K Urakami, and M Shoji. Angiotensin-converting enzyme genotype is associated

with Alzheimer disease in the Japanese population. Neurosci Lett, 277(1):65–7,

1999.

H-M Huang, Y-M Kuo, H-C Ou, C-C Lin, and L-J Chuo. Apolipoprotein E polymor-

phism in various dementias in Taiwan Chinese population. J Neural Transm, 109

(11):1415–21, 2002.

BIBLIOGRAPHY 107

I Ikaheimo, S Silvennoinen-Kassinen, J Karvonen, T Jarvinen, and A Ti-

ilikainen. Immunogenetic profile of psoriasis vulgaris: association with haplotypes

A2,B13,Cw6,DR7,DQA1*0201 and A1,B17,Cw6,DR7,DQA1*0201. Arch Dermatol

Res, 288(2):63–7, 1996.

T Isbir, B Agachan, H Yilmaz, and M Aydin. Angiotensin converting enzyme gene

polymorphism in Alzheimer’s disease. Cell Biochem Funct, 18(2):141–2, 2000.

M Jobim, L F J Jobim, P H Salim, T F Cestari, R Toresan, B C Gil, M R Jobim, T J

Wilson, M Kruger, J Schlottfeldt, and G Schwartsmann. A study of the killer cell

immunoglobulin-like receptor gene KIR2DS1 in a Caucasoid Brazilian population

with psoriasis vulgaris. Tissue Antigens, 72(4):392–6, 2008.

Anna Juhasz, Agnes Rimanoczy, Krisztina Boda, Gabor Vincze, Gyozo Szlavik, Mari-

anna Zana, Annamaria Bjelik, Magdolna Pakaski, Nikoletta Bodi, Andras Palotas,

Zoltan Janka, and Janos Kalman. CYP46 T/C polymorphism is not associated

with Alzheimer’s dementia in a population from Hungary. Neurochem Res, 30(8):

943–8, 2005.

M Kastelan, F Gruber, E Cecuk, V Kerhin-Brkljacic, L Brkljacic-Surkalovic, and

A Kastelan. Analysis of HLA antigens in Croatian patients with psoriasis. Acta

Derm Venereol Suppl (Stockh), NIL(211):12–3, 2000.

H C Kim, D K Kim, I J Choi, K H Kang, S D Yi, J Park, and Y N Park. Relation of

apolipoprotein E polymorphism to clinically diagnosed Alzheimer’s disease in the

Korean population. Psychiatry Clin Neurosci, 55(2):115–20, 2001.

K W Kim, J H Jhoo, K U Lee, D Y Lee, J H Lee, J Y Youn, B J Lee, S H Han,

and J I Woo. No association between alpha-1-antichymotrypsin polymorphism and

Alzheimer’s disease in Koreans. Am J Med Genet, 91(5):355–8, 2000a.

T G Kim, H J Lee, J I Youn, T Y Kim, and H Han. The association of psoriasis with

human leukocyte antigens in Korean population and the influence of age of onset

and sex. J Invest Dermatol, 114(2):309–13, 2000b.

108 BIBLIOGRAPHY

H Kolsch, F Jessen, N Freymann, M Kreis, F Hentschel, W Maier, and R Heun.

ACE I/D polymorphism is a risk factor of Alzheimer’s disease but not of vascular

dementia. Neurosci Lett, 377(1):37–9, 2005.

W A Kukull, G D Schellenberg, J D Bowen, W C McCormick, C E Yu, L Teri,

J D Thompson, E S O’Meara, and E B Larson. Apolipoprotein E in Alzheimer’s

disease risk and case detection: a case-control study. J Clin Epidemiol, 49(10):

1143–8, 1996.

S. Kullback and R. A. Leibler. On information and sufficiency. The An-

nals of Mathematical Statistics, 22(1):79–86, 1951. ISSN 00034851. URL

http://www.jstor.org/stable/2236703.

N Kundakci, T Oskay, U Olmez, H Tutkak, and E Gurgey. Association of psoriasis

vulgaris with HLA class I and class II antigens in the Turkish population, according

to the age at onset. Int J Dermatol, 41(6):345–8, 2002.

H Kunugi, A Akahane, A Ueki, M Otsuka, K Isse, H Hirasawa, N Kato, T Nabika,

S Kobayashi, and S Nanko. No evidence for an association between the Glu298Asp

polymorphism of the NOS3 gene and Alzheimer’s disease. J Neural Transm, 107

(8-9):1081–4, 2000.

J C Lambert, C Berr, F Pasquier, A Delacourte, B Frigard, D Cottel, J Perez-Tur,

V Mouroux, M Mohr, D Cecyre, D Galasko, C Lendon, J Poirier, J Hardy, D Mann,

P Amouyel, and M C Chartier-Harlin. Pronounced impact of Th1/E47cs mutation

compared with -491 AT mutation on neural APOE gene expression and risk of

developing Alzheimer’s disease. Hum Mol Genet, 7(9):1511–6, 1998.

L Lannfelt, L Lilius, M Nastase, M Viitanen, L Fratiglioni, G Eggertsen, L Berglund,

B Angelin, J Linder, and B Winblad. Lack of association between apolipoprotein

E allele epsilon 4 and sporadic Alzheimer’s disease. Neurosci Lett, 169(1-2):175–8,

1994.

E.L. Lehmann and G. Casella. Theory of Point Estimation. Springer Verlag, New

York, 2nd edition, 1998.

BIBLIOGRAPHY 109

M Lehtovirta, H Soininen, S Helisalmi, A Mannermaa, E L Helkala, P Hartikainen,

T Hanninen, M Ryynanen, and P J Riekkinen. Clinical and neuropsychological

characteristics in familial and sporadic Alzheimer’s disease: relation to apolipopro-

tein E polymorphism. Neurology, 46(2):413–9, 1996.

C L Lendon, U Thaker, J M Harris, A M McDonagh, J-C Lambert, M-C Chartier-

Harlin, T Iwatsubo, S M Pickering-Brown, and D M A Mann. The angiotensin 1-

converting enzyme insertion (I)/deletion (D) polymorphism does not influence the

extent of amyloid or tau pathology in patients with sporadic Alzheimer’s disease.

Neurosci Lett, 328(3):314–8, 2002.

Hsien-Tzung Liao, Kuan-Chia Lin, Yun-Ting Chang, Chun-Hsiung Chen, Toong-Hua

Liang, Wei-Sheng Chen, Kuei-Ying Su, Chang-Youh Tsai, and Chung-Tei Chou.

Human leukocyte antigen and clinical and demographic characteristics in psoriatic

arthritis and psoriasis in Chinese patients. J Rheumatol, 35(5):891–5, 2008.

Roderick J.A. Little and Donald B. Rubin. Statistical Analysis with Missing Data.

Wiley and Sons, New York, 2nd edition, 2002.

H C Liu, C J Hong, S J Wang, J L Fuh, P N Wang, H Y Shyu, and E L Teng.

ApoE genotype in relation to AD and cholesterol: a study of 2,326 Chinese adults.

Neurology, 53(5):962–6, 1999.

O L Lopez, S Lopez-Pousa, M I Kamboh, R Adroer, R Oliva, M Lozano-Gallego, J T

Becker, and S T DeKosky. Apolipoprotein E polymorphism in Alzheimer’s disease:

a comparative study of two research populations from Spain and the United States.

Eur Neurol, 39(4):229–33, 1998.

C Lopez-Larrea, J C Torre Alonso, A Rodriguez Perez, and E Coto. HLA antigens in

psoriatic arthritis subtypes of a Spanish population. Ann Rheum Dis, 49(5):318–9,

1990.

Thomas A. Louis. Finding the observed information matrix when using the em al-

gorithm. Journal of the Royal Statistical Society. Series B (Methodological), 44(2):

226–233, 1982. ISSN 00359246. URL http://www.jstor.org/stable/2345828.

110 BIBLIOGRAPHY

Wioleta Luszczek, Wioletta Kubicka, Maria Cislo, Piotr Nockowski, Maria Manczak,

Grzegorz Woszczek, Eugeniusz Baran, and Piotr Kusnierczyk. Strong association

of HLA-Cw6 allele with juvenile psoriasis in Polish patients. Immunol Lett, 85(1):

59–64, 2003.

Suk Ling Ma, Nelson Leung Sang Tang, Linda Chiu Wa Lam, and Helen Fung Kum

Chiu. The association between promoter polymorphism of the interleukin-10 gene

and Alzheimer’s disease. Neurobiol Aging, 26(7):1005–10, 2005.

G Maestre, R Ottman, Y Stern, B Gurland, M Chun, M X Tang, M Shelanski, B Ty-

cko, and R Mayeux. Apolipoprotein E and Alzheimer’s disease: ethnic variation in

genotypic risks. Ann Neurol, 37(2):254–9, 1995.

E Mallon, M Bunce, F Wojnarowska, and K Welsh. HLA-CW*0602 is a susceptibility

factor in type I psoriasis, and evidence Ala-73 is increased in male type I psoriatics.

J Invest Dermatol, 109(2):183–6, 1997.

E Mallon, D Young, M Bunce, F M Gotch, P J Easterbrook, R Newson, and C B

Bunker. HLA-Cw*0602 and HIV-associated psoriasis. Br J Dermatol, 139(3):527–

33, 1998.

E Mallon, M Bunce, H Savoie, A Rowe, R Newson, F Gotch, and C B Bunker. HLA-C

and guttate psoriasis. Br J Dermatol, 143(6):1177–82, 2000.

N. Mantel and W. Haenszel. Statistical aspects of the analysis of data from retro-

spective studies of disease. J. Natl. Cancer Inst., 22:719–748, 1959.

Maureen P Martin, George Nelson, Jeong-Hee Lee, Fawnda Pellett, Xiaojiang Gao,

Judith Wade, Michael J Wilson, John Trowsdale, Dafna Gladman, and Mary Car-

rington. Cutting edge: susceptibility to psoriatic arthritis: influence of activating

killer Ig-like receptor genes in the absence of specific HLA-C alleles. J Immunol,

169(6):2818–22, 2002.

J Martinez-Borra, S Gonzalez, J Santos-Juanes, J Sanchez del Rio, J C Torre-Alonso,

A Lopez-Vazquez, M A Blanco-Gelaz, and C Lopez-Larrea. Psoriasis vulgaris and

BIBLIOGRAPHY 111

psoriatic arthritis share a 100 kb susceptibility region telomeric to HLA-C. Rheuma-

tology (Oxford), 42(9):1089–92, 2003.

P. McCullagh and J.A. Nelder. Generalized Linear Models. Chapman and Hall,

London, 2nd edition, 1983.

F O Meenan. A note on the history of psoriasis. Ir J Med Sci, 6(351):141–2, 1955.

Xiao-Li Meng and Donald B. Rubin. Using em to obtain asymptotic

variance-covariance matrices: The sem algorithm. Journal of the Ameri-

can Statistical Association, 86(416):899–909, 1991. ISSN 01621459. URL

http://www.jstor.org/stable/2290503.

X.L. Meng and D.B. Rubin. Maximum likelihood estimation via the ecm algorithm:

A general framework. Biometrika, 80:267–278, 1993.

A E Molero, G Pino-Ramirez, and G E Maestre. Modulation by age and gender of risk

for Alzheimer’s disease and vascular dementia associated with the apolipoprotein

E-epsilon4 allele in Latin Americans: findings from the Maracaibo Aging Study.

Neurosci Lett, 307(1):5–8, 2001.

Roberto Monastero, Rosalia Caldarella, Marina Mannino, Angelo B Cefalu, Gianluca

Lopez, Davide Noto, Cecilia Camarda, Lawrence K C Camarda, Alberto Notar-

bartolo, Maurizio R Averna, and Rosolino Camarda. Lack of association between

angiotensin converting enzyme polymorphism and sporadic Alzheimer’s disease.

Neurosci Lett, 335(2):147–9, 2002.

Roberto Monastero, Angelo B Cefalu, Cecilia Camarda, Carmela M Buglino, Marina

Mannino, Carlo M Barbagallo, Gianluca Lopez, Lawrence K C Camarda, Salva-

tore Travali, Rosolino Camarda, and Maurizio R Averna. No association between

Glu298Asp endothelial nitric oxide synthase polymorphism and Italian sporadic

Alzheimer’s disease. Neurosci Lett, 341(3):229–32, 2003.

S Mui, M Briggs, H Chung, R B Wallace, T Gomez-Isla, G W Rebeck, and B T

Hyman. A newly identified polymorphism in the apolipoprotein E enhancer gene

112 BIBLIOGRAPHY

region is associated with Alzheimer’s disease and strongly with the epsilon 4 allele.

Neurology, 47(1):196–201, 1996.

C Murray, D L Mann, L N Gerber, W Barth, S Perlmann, J L Decker, and T P

Nigra. Histocompatibility alloantigens in psoriasis and psoriatic arthritis. Evidence

for the influence of multiple genes in the major histocompatibility complex. J Clin

Invest, 66(4):670–5, 1980.

L Myllykangas, T Polvikoski, R Sulkava, A Verkkoniemi, P Tienari, L Niinisto,

K Kontula, J Hardy, M Haltia, and J Perez-Tur. Cardiovascular risk factors and

Alzheimer’s disease: a genetic association study in a population aged 85 or over.

Neurosci Lett, 292(3):195–8, 2000.

Rajan P Nair, Andreas Ruether, Philip E Stuart, Stefan Jenisch, Trilokraj Tejasvi,

Ravi Hiremagalore, Stefan Schreiber, Dieter Kabelitz, Henry W Lim, John J

Voorhees, Enno Christophers, James T Elder, and Michael Weichenthal. Poly-

morphisms of the IL12B and IL23R genes are associated with psoriasis. J Invest

Dermatol, 128(7):1653–61, 2008.

H Nakagawa, S Akazaki, A Asahina, K Tokunaga, K Matsuki, S Kuwata, Y Ishibashi,

and T Juji. Study of HLA class I, class II and complement genes (C2, C4A, C4B

and BF) in Japanese psoriatics and analysis of a newly-found high-risk haplotype

by pulsed field gel electrophoresis. Arch Dermatol Res, 283(5):281–4, 1991.

S Nakayama and S Kuzuhara. Apolipoprotein E phenotypes in healthy normal con-

trols and demented subjects with Alzheimer’s disease and vascular dementia in Mie

Prefecture of Japan. Psychiatry Clin Neurosci, 53(6):643–8, 1999.

J Nalbantoglu, B M Gilfix, P Bertrand, Y Robitaille, S Gauthier, D S Rosenblatt, and

J Poirier. Predictive value of apolipoprotein E genotyping in Alzheimer’s disease:

results of an autopsy series and an analysis of several combined studies. Ann Neurol,

36(6):889–95, 1994.

Y Narain, A Yip, T Murphy, C Brayne, D Easton, J G Evans, J Xuereb, N Cairns,

BIBLIOGRAPHY 113

M M Esiri, R A Furlong, and D C Rubinsztein. The ACE gene and Alzheimer’s

disease susceptibility. J Med Genet, 37(9):695–7, 2000.

Frank O Nestle, Daniel H Kaplan, and Jonathan Barker. Psoriasis. N Engl J Med,

361(5):496–509, 2009.

A Nunomura, S Chiba, M Eto, M Saito, I Makino, and T Miyagishi. Apolipoprotein

E polymorphism and susceptibility to early- and late-onset sporadic Alzheimer’s

disease in Hokkaido, the northern part of Japan. Neurosci Lett, 206(1):17–20, 1996.

K P O’Brien, S J Holm, S Nilsson, L Carlen, T Rosenmuller, C Enerback, A Inerot,

and M Stahle-Backdahl. The HCR gene on 6p21 is unlikely to be a psoriasis

susceptibility gene. J Invest Dermatol, 116(5):750–4, 2001.

S Orru, E Giuressi, M Casula, A Loizedda, R Murru, M Mulargia, M V Masala,

D Cerimele, M Zucca, N Aste, P Biggio, C Carcassi, and L Contu. Psoriasis is

associated with a SNP haplotype of the corneodesmosin gene (CDSN). Tissue

Antigens, 60(4):292–8, 2002.

B O Osuntokun, A Sahota, A O Ogunniyi, O Gureje, O Baiyewu, A Adeyinka, S O

Oluwole, O Komolafe, K S Hall, and F W Unverzagt. Lack of an association

between apolipoprotein E epsilon 4 and Alzheimer’s disease in elderly Nigerians.

Ann Neurol, 38(3):463–5, 1995.

A Ozawa, M Ohkido, H Inoko, A Ando, and K Tsuji. Specific restriction fragment

length polymorphism on the HLA-C region and susceptibility to psoriasis vulgaris.

J Invest Dermatol, 90(3):402–5, 1988.

F Panza, V Solfrizzi, F Torres, F Mastroianni, A M Colacicco, A M Basile, C Capurso,

A D’Introno, A Del Parigi, and A Capurso. Apolipoprotein E in Southern Italy:

protective effect of epsilon 2 allele in early- and late-onset sporadic Alzheimer’s

disease. Neurosci Lett, 292(2):79–82, 2000.

Francesco Panza, Vincenzo Solfrizzi, Anna M Colacicco, Anna M Basile, Alessia

114 BIBLIOGRAPHY

D’Introno, Cristiano Capurso, Maria Sabba, Sabrina Capurso, and Antonio Ca-

purso. Apolipoprotein E (APOE) polymorphism influences serum APOE levels in

Alzheimer’s disease patients and centenarians. Neuroreport, 14(4):605–8, 2003.

Yudi Pawitan. In All Likelihood. Oxford University Press, Oxford, 2001.

R T Perry, J S Collins, L E Harrell, R T Acton, and R C Go. Investigation of

association of 13 polymorphisms in eight genes in southeastern African American

Alzheimer disease patients as compared to age-matched controls. Am J Med Genet,

105(4):332–42, 2001.

J Poirier, J Davignon, D Bouthillier, S Kogan, P Bertrand, and S Gauthier.

Apolipoprotein E polymorphism and Alzheimer’s disease. Lancet, 342(8873):697–9,

1993.

J A Prince, L Feuk, S L Sawyer, J Gottfries, A Ricksten, K Nagga, N Bogdanovic,

K Blennow, and A J Brookes. Lack of replication of association findings in complex

disease: an analysis of 15 polymorphisms in prior candidate genes for sporadic

Alzheimer’s disease. Eur J Hum Genet, 9(6):437–44, 2001.

Chul-Woo Pyo, Seong-Suk Hur, Yang-Kyum Kim, Tae-Yoon Kim, and Tai-Gyu Kim.

Association of TAP and HLA-DM genes with psoriasis in Koreans. J Invest Der-

matol, 120(4):616–22, 2003.

R Queiro, P Moreno, C Sarasqueta, M Alperi, J L Riestra, and J Ballina. Synovitis-

acne-pustulosis-hyperostosis-osteitis syndrome and psoriatic arthritis exhibit a dif-

ferent immunogenetic profile. Clin Exp Rheumatol, 26(1):125–8, 2008.

Ruben Queiro, Juan Carlos Torre, Segundo Gonzalez, Carlos Lopez-Larrea, Tomas

Tinture, and Isaac Lopez-Lagunas. HLA antigens may influence the age of onset

of psoriasis and psoriatic arthritis. J Rheumatol, 30(3):505–7, 2003.

Ruben Queiro, Segundo Gonzalez, Carlos Lopez-Larrea, Mercedes Alperi, Cristina

Sarasqueta, Jose Luis Riestra, and Javier Ballina. HLA-C locus alleles may mod-

ulate the clinical expression of psoriatic arthritis. Arthritis Res Ther, 8(6):R185,

2006.

BIBLIOGRAPHY 115

R Queiro-Silva, J C Torre-Alonso, T Tinture-Eguren, and I Lopez-Lagunas. The effect

of HLA-DR antigens on the susceptibility to, and clinical expression of psoriatic

arthritis. Scand J Rheumatol, 33(5):318–22, 2004.

P Quiroga, C Calvo, C Albala, J Urquidi, J L Santos, H Perez, and G Klaassen.

Apolipoprotein E polymorphism in elderly Chilean people with Alzheimer’s disease.

Neuroepidemiology, 18(1):48–52, 1999.

P Rahman, S Bartlett, F Siannis, F J Pellett, V T Farewell, L Peddle, C T Schentag,

C A Alderdice, S Hamilton, M Khraishi, Y Tobin, D Hefferton, and D D Gladman.

CARD15: a pleiotropic autoimmune gene that confers susceptibility to psoriatic

arthritis. Am J Hum Genet, 73(3):677–81, 2003.

R Rani, R Narayan, M A Fernandez-Vina, and P Stastny. Role of HLA-B and C

alleles in development of psoriasis in patients from North India. Tissue Antigens,

51(6):618–22, 1998.

Asad Vaisi Raygani, Mahine Zahrai, Akbar Vaisi Raygani, Mahmood Doosti, Ebrahim

Javadi, Mansour Rezaei, and Tayebeh Pourmotabbed. Association between

apolipoprotein E polymorphism and Alzheimer disease in Tehran, Iran. Neurosci

Lett, 375(1):1–6, 2005.

J Reefhuis, M A Honein, L A Schieve, A Correa, C A Hobbs, and S A Rasmussen.

Assisted reproductive technology and major structural birth defects in the United

States. Hum Reprod, 24(2):360–6, 2009.

F Richard, I Fromentin-David, F Ricolfi, P Ducimetiere, C Di Menza, P Amouyel,

and N Helbecque. The angiotensin I converting enzyme gene as a susceptibility

factor for dementia. Neurology, 56(11):1593–5, 2001.

James Robins, Norman Breslow, and Sander Greenland. Estimators of the

mantel-haenszel variance consistent in both sparse data and large-strata lim-

iting models. Biometrics, 42(2):311–323, 1986. ISSN 0006341X. URL

http://www.jstor.org/stable/2531052.

116 BIBLIOGRAPHY

A Roitberg-Tambur, A Friedmann, E E Tzfoni, S Battat, R Ben Hammo, C Safirman,

K Tokunaga, A Asahina, and C Brautbar. Do specific pockets of HLA-C molecules

predispose Jewish patients to psoriasis vulgaris? J Am Acad Dermatol, 31(6):

964–8, 1994.

Stavra N Romas, Vincent Santana, Jennifer Williamson, Alejandra Ciappa, Joseph H

Lee, Haydee Z Rondon, Pedro Estevez, Rafael Lantigua, Martin Medrano, May-

obanex Torres, Yaakov Stern, Benjamin Tycko, and Richard Mayeux. Familial

Alzheimer disease among Caribbean Hispanics: a reexamination of its association

with APOE. Arch Neurol, 59(1):87–91, 2002.

A V Romphruk, A Oka, A Romphruk, M Tomizawa, C Choonhakarn, T K Naruse,

C Puapairoj, G Tamiya, C Leelayuwat, and H Inoko. Corneodesmosin gene: no ev-

idence for PSORS 1 gene in North-eastern Thai psoriasis patients. Tissue Antigens,

62(3):217–24, 2003.

A D Roses. A model for susceptibility polymorphisms for complex diseases:

apolipoprotein E and Alzheimer disease. Neurogenetics, 1(1):3–11, 1997.

T J Russell, L M Schultes, and D J Kuban. Histocompatibility (HL-A) antigens

associated with psoriasis. N Engl J Med, 287(15):738–40, 1972.

A Sahota, M Yang, S Gao, S L Hui, O Baiyewu, O Gureje, S Oluwole, A Ogunniyi,

K S Hall, and H C Hendrie. Apolipoprotein E-associated risk for Alzheimer’s

disease in the African-American population is genotype dependent. Ann Neurol,

42(4):659–61, 1997.

Fabio Sanchez, Sofia J Holm, Lotus Mallbris, Kevin P O’Brien, and Mona Stahle.

STG does not associate with psoriasis in the Swedish population. Exp Dermatol,

13(7):413–8, 2004.

Fabio O Sanchez, M V Prasad Linga Reddy, Lotus Mallbris, Kazuko Sakuraba, Mona

Stahle, and Marta E Alarcon-Riquelme. IFN-regulatory factor 5 gene variants

interact with the class I MHC locus in the Swedish psoriasis population. J Invest

Dermatol, 128(7):1704–9, 2008.

BIBLIOGRAPHY 117

M Sanchez-Guerra, O Combarros, A Alvarez-Arcaya, I Mateo, J Berciano,

J Gonzalez-Garcia, and J Llorca. The Glu298Asp polymorphism in the NOS3 gene

is not associated with sporadic Alzheimer’s disease. J Neurol Neurosurg Psychiatry,

70(4):566–7, 2001.

R Scacchi, L De Bernardini, E Mantuano, L M Donini, T Vilardo, and R M Corbo.

Apolipoprotein E (APOE) allele frequencies in late-onset sporadic Alzheimer’s dis-

ease (AD), mixed dementia and vascular dementia: lack of association of epsilon 4

allele with AD in Italian octogenarian patients. Neurosci Lett, 201(3):231–4, 1995.

Joeseph L. Schafer. Analysis of Incomplete Multivariate Data. Chapman and Hall,

New York, 1997.

M Schmitt-Egenolf, T H Eiermann, W H Boehncke, M Stander, and W Sterry.

Familial juvenile onset psoriasis is associated with the human leukocyte antigen

(HLA) class I side of the extended haplotype Cw6-B57-DRB1*0701-DQA1*0201-

DQB1*0303: a population- and family-based study. J Invest Dermatol, 106(4):

711–4, 1996.

W K Scott, A M Saunders, P C Gaskell, P A Locke, J H Growdon, L A Farrer,

S A Auerbach, A D Roses, J L Haines, and M A Pericak-Vance. Apolipoprotein

E epsilon2 does not increase risk of early-onset sporadic Alzheimer’s disease. Ann

Neurol, 42(3):376–8, 1997.

Shaun R. Seaman and Sylvia Richardson. Bayesian analysis of case-control studies

with categorical covariates. Biometrika, 88(4):1073–1088, 2001. ISSN 00063444.

URL http://www.jstor.org/stable/2673702.

D Seripa, M G Matera, R P D’Andrea, C Gravina, C Masullo, A Daniele, A Bizzarro,

M Rinaldi, P Antuono, D R Wekstein, G Dal Forno, and V M Fazio. Alzheimer

disease risk associated with APOE4 is modified by STH gene polymorphism. Neu-

rology, 62(9):1631–3, 2004.

Davide Seripa, Gloria Dal Forno, Maria G Matera, Carolina Gravina, Maurizio Mar-

gaglione, Mark T Palermo, David R Wekstein, Piero Antuono, Daron G Davis,

118 BIBLIOGRAPHY

Antonio Daniele, Carlo Masullo, Alessandra Bizzarro, Massimo Gennarelli, and

Vito M Fazio. Methylenetetrahydrofolate reductase and angiotensin converting en-

zyme gene polymorphisms in two genetically and diagnostically distinct cohort of

Alzheimer patients. Neurobiol Aging, 24(7):933–9, 2003.

A B Singleton, A M Gibson, I G McKeith, C G Ballard, J A Edwardson, and C M

Morris. Nitric oxide synthase gene polymorphisms in Alzheimer’s disease and de-

mentia with Lewy bodies. Neurosci Lett, 303(1):33–6, 2001.

Kristel Sleegers, Tom den Heijer, Ewoud J van Dijk, Albert Hofman, Aida M Bertoli-

Avella, Peter J Koudstaal, Monique M B Breteler, and Cornelia M van Duijn.

ACE gene is associated with Alzheimer’s disease and atrophy of hippocampus and

amygdala. Neurobiol Aging, 26(8):1153–9, 2005.

A J Slooter, M Cruts, S Kalmijn, A Hofman, M M Breteler, C Van Broeckhoven,

and C M van Duijn. Risk estimates of dementia by apolipoprotein E genotypes

from a population-based incidence study: the Rotterdam Study. Arch Neurol, 55

(7):964–8, 1998.

S Sorbi, B Nacmias, P Forleo, S Latorraca, I Gobbini, L Bracco, S Piacentini, and

L Amaducci. ApoE allele frequencies in Italian sporadic and familial Alzheimer’s

disease. Neurosci Lett, 177(1-2):100–2, 1994.

D R S Souza, M R de Godoy, J Hotta, E H Tajara, A C Brandao, S Pinheiro Junior,

W A Tognola, and J E dos Santos. Association of apolipoprotein E polymorphism

in late-onset Alzheimer’s disease and vascular dementia in Brazilians. Braz J Med

Biol Res, 36(7):919–23, 2003.

D F Stroup, J A Berlin, S C Morton, I Olkin, G D Williamson, D Rennie, D Moher,

B J Becker, T A Sipe, and S B Thacker. Meta-analysis of observational studies in

epidemiology: a proposal for reporting. Meta-analysis Of Observational Studies in

Epidemiology (MOOSE) group. JAMA, 283(15):2008–12, 2000.

Maria Styczynska, Dorota Religa, Anna Pfeffer, Elzbieta Luczywek, Boguslaw

Wasiak, Grzegorz Styczynski, Beata Peplonska, Tomasz Gabryelewicz, Marek

BIBLIOGRAPHY 119

Golebiowski, Malgorzata Kobrys, and Maria Barcikowska. Simultaneous analy-

sis of five genetic risk factors in Polish patients with Alzheimer’s disease. Neurosci

Lett, 344(2):99–102, 2003.

R Sulkava, K Kainulainen, A Verkkoniemi, L Niinisto, E Sobel, Z Davanipour,

T Polvikoski, M Haltia, and K Kontula. APOE alleles in Alzheimer’s disease and

vascular dementia in a population aged 85+. Neurobiol Aging, 17(3):373–6, 1996.

Trey Sunderland, Nadeem Mirza, Karen T Putnam, Gary Linker, Deepa Bhupali,

Rob Durham, Holly Soares, Lida Kimmel, David Friedman, Judy Bergeson, Gyorgy

Csako, James A Levy, John J Bartko, and Robert M Cohen. Cerebrospinal fluid

beta-amyloid1-42 and tau in control subjects at risk for Alzheimer’s disease: the

effect of APOE epsilon4 allele. Biol Psychiatry, 56(9):670–6, 2004.

Aneta Szczerkowska-Dobosz, Krzysztof Rebala, Zofia Szczerkowska, and Anna

Witkowska-Tobola. Correlation of HLA-Cw*06 allele frequency with some clin-

ical features of psoriasis vulgaris in the population of northern Poland. J Appl

Genet, 45(4):473–6, 2004.

C Talbot, C Lendon, N Craddock, S Shears, J C Morris, and A Goate. Protection

against Alzheimer’s disease with apoE epsilon 2. Lancet, 343(8910):1432–3, 1994.

M X Tang, Y Stern, K Marder, K Bell, B Gurland, R Lantigua, H Andrews, L Feng,

B Tycko, and R Mayeux. The APOE-epsilon4 allele and the risk of Alzheimer

disease among African Americans, whites, and Hispanics. JAMA, 279(10):751–5,

1998.

Martin A. Tanner and Wing Hung Wong. The calculation of posterior distributions

by data augmentation. Journal of the American Statistical Association, 82(398):

528–540, 1987. ISSN 01621459. URL http://www.jstor.org/stable/2289457.

T Tapiola, M Lehtovirta, J Ramberg, S Helisalmi, K Linnaranta, P Sr Riekkinen, and

H Soininen. CSF tau is related to apolipoprotein E genotype in early Alzheimer’s

disease. Neurology, 50(1):169–74, 1998.

120 BIBLIOGRAPHY

Andrea Tedde, Benedetta Nacmias, Elena Cellini, Silvia Bagnoli, and Sandro Sorbi.

Lack of association between NOS3 poly morphism and Italian sporadic and familial

Alzheimer’s disease. J Neurol, 249(1):110–1, 2002.

L Tilley, K Morgan, J Grainger, P Marsters, L Morgan, J Lowe, J Xuereb, C Wischik,

C Harrington, and N Kalsheker. Evaluation of polymorphisms in the presenilin-

1 gene and the butyrylcholinesterase gene as risk factors in sporadic Alzheimer’s

disease. Eur J Hum Genet, 7(6):659–63, 1999.

T Town, D Paris, D Fallin, R Duara, W Barker, M Gold, F Crawford, and M Mul-

lan. The -491A/T apolipoprotein E promoter polymorphism association with

Alzheimer’s disease: independent risk and linkage disequilibrium with the known

APOE polymorphism. Neurosci Lett, 252(2):95–8, 1998.

D W Tsuang, R K Wilson, O L Lopez, E K Luedecking-Zimmer, J B Leverenz,

S T DeKosky, M I Kamboh, and R L Hamilton. Genetic association between the

APOE*4 allele and Lewy bodies in Alzheimer disease. Neurology, 64(3):509–13,

2005.

C M van Duijn, P de Knijff, A Wehnert, J De Voecht, J B Bronzova, L M Havekes,

A Hofman, and C Van Broeckhoven. The apolipoprotein E epsilon 2 allele is

associated with an increased risk of early-onset Alzheimer’s disease and a reduced

survival. Ann Neurol, 37(5):605–10, 1995.

S Vejbaesya, T H Eiermann, P Suthipinititharm, C Bancha, H A Stephens, K Lu-

angtrakool, and D Chandanayingyong. Serological and molecular analysis of HLA

class I and II alleles in Thai patients with psoriasis vulgaris. Tissue Antigens, 52

(4):389–92, 1998.

Simona Vuletic, Elaine R Peskind, Santica M Marcovina, Joseph F Quinn, Marian C

Cheung, Hal Kennedy, Jeffrey A Kaye, Lee-Way Jin, and John J Albers. Reduced

CSF PLTP activity in Alzheimer’s disease and other neurologic diseases; PLTP

induces ApoE secretion in primary human astrocytes in vitro. J Neurosci Res, 80

(3):406–13, 2005.

BIBLIOGRAPHY 121

Yosuke Wakutani, Hisanori Kowa, Masayoshi Kusumi, Kaoru Yamagata, Kenji Wada-

Isoe, Yoshiki Adachi, Takao Takeshima, Katsuya Urakami, and Kenji Nakashima.

Genetic analysis of vascular factors in Alzheimer’s disease. Ann N Y Acad Sci, 977

(NIL):232–8, 2002.

H K Wang, H C Fung, W C Hsu, Y R Wu, J C Lin, L S Ro, K H Chang, F J Hwu,

Y Hsu, S Y Huang, G J Lee-Chen, and C M Chen. Apolipoprotein E, angiotensin-

converting enzyme and kallikrein gene polymorphisms and the risk of Alzheimer’s

disease and vascular dementia. J Neural Transm, 113(10):1499–509, 2006.

J C Wang, J M Kwon, P Shah, J C Morris, and A Goate. Effect of APOE genotype

and promoter polymorphism on risk of Alzheimer’s disease. Neurology, 55(11):

1644–9, 2000.

Stephen C Waring and Roger N Rosenberg. Genome-wide association studies in

Alzheimer disease. Arch Neurol, 65(3):329–34, 2008.

H Wiebusch, J Poirier, P Sevigny, and K Schappert. Further evidence for a synergistic

association between APOE epsilon4 and BCHE-K in confirmed Alzheimer’s disease.

Hum Genet, 104(2):158–63, 1999.

Fionnuala Williams, Ashley Meenagh, Carole Sleator, Daniel Cook, Marcelo

Fernandez-Vina, Anne M Bowcock, and Derek Middleton. Activating killer cell

immunoglobulin-like receptor gene KIR2DS1 is associated with psoriatic arthritis.

Hum Immunol, 66(7):836–41, 2005.

Andrzej Wisniewski, Wioleta Luszczek, Maria Manczak, Monika Jasek, Wioletta Ku-

bicka, Maria Cislo, and Piotr Kusnierczyk. Distribution of LILRA3 (ILT6/LIR4)

deletion in psoriatic patients and healthy controls. Hum Immunol, 64(4):458–61,

2003.

J D Yang, G Feng, J Zhang, Z X Lin, T Shen, G Breen, D St Clair, and L He. As-

sociation between angiotensin-converting enzyme gene and late onset Alzheimer’s

disease in Han chinese. Neurosci Lett, 295(1-2):41–4, 2000.

122 BIBLIOGRAPHY

J D Yang, G Y Feng, J Zhang, J Cheung, D St Clair, L He, and Keiichi Ichimura.

Apolipoprotein E -491 promoter polymorphism is an independent risk factor for

Alzheimer’s disease in the Chinese population. Neurosci Lett, 350(1):25–8, 2003.

Pamela Zambenedetti, GianLuca De Bellis, Ida Biunno, Massimo Musicco, and Paolo

Zatta. Transferrin C2 variant does confer a risk for Alzheimer’s disease in cau-

casians. J Alzheimers Dis, 5(6):423–7, 2003.

Peng Zhang, Ze Yang, Chuanfang Zhang, Zeping Lu, Xiaohong Shi, Weidong Zheng,

Chunling Wan, Duanyang Zhang, Chenguang Zheng, Shu Li, Feng Jin, and

Li Wang. Association study between late-onset Alzheimer’s disease and the trans-

ferrin gene polymorphisms in Chinese. Neurosci Lett, 349(3):209–11, 2003.

G Zuliani, A Ble’, R Zanca, M R Munari, A Zurlo, C Vavalle, A R Atti, and R Fellin.

Genetic polymorphisms in older subjects with vascular or Alzheimer’s dementia.

Acta Neurol Scand, 103(5):304–8, 2001.