An Introduction to Veterinary Epidemiology

Mark Stevenson
EpiCentre, IVABS, Massey University, Palmerston North
New Zealand
 An Introduction to Veterinary Epidemiology ∗

Mark Stevenson †
EpiCentre, IVABS ‡
Massey University, Palmerston North, New Zealand

August 13, 2012

Contents

1 Introduction 6
1.1 Definitions and objectives of epidemiology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Host, agent, and environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Individual, place, and time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Individual . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Place . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Measures of health 13
2.1 Prevalence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Incidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Incidence risk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
Incidence rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
The relationship between prevalence and incidence . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Other measures of health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Attack rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Secondary attack rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Crude mortality rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Age-specific mortality rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
∗
Notes for Veterinary Biometrics and Epidemiology, taught as course 227.407 within the BVSc program at Massey
University. Contributions from Dirk Pfeiffer, Cord Heuer, Nigel Perkins, and John Morton are gratefully acknowledged.
†
[email protected]
‡
URL: http://epicentre.massey.ac.nz

1
 Case fatality rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Proportional mortality rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.4 Adjusted measures of health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
Direct adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
Indirect adjustment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3 Study design 26
3.1 Descriptive studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Case reports . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Cases series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Descriptive studies based on rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 Analytical studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Ecological studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Cross-sectional studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Cohort studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Case-control studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
Hybrid study designs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Randomised clinical trials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
Community trials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.3 Comparison of the major epidemiological study designs . . . . . . . . . . . . . . . . . . . . . 35

4 Measures of association 37
4.1 Measures of strength . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Incidence risk ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Incidence rate ratio . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
Odds ratio — cohort studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Odds ratio — case-control studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.2 Measures of effect in the exposed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Attributable risk (rate) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
Attributable fraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3 Measures of effect in the population . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Population attributable risk (rate) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
Population attributable fraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
4.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

2
5 Error in epidemiological research 46
5.1 Accuracy and precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
5.2 Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Selection bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
Misclassification bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.3 Confounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.4 Interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
5.5 Dealing with confounding and interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
Methods for dealing with confounding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
A worked example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

6 Causation 63
6.1 Association versus causation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
6.2 Component, sufficient, and necessary causes . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
6.3 Hill’s criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
Strength of association . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Temporality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Dose response . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
Plausibility and coherence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Experimental evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Specificity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Analogy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
6.4 Causal web models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68

7 Sampling 70
7.1 Probability sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Simple random sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Systematic random sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Stratified random sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
Cluster sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
7.2 Non-probability sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
7.3 Sampling techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
Methods of randomisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Replacement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Probability proportional to size sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
7.4 Sample size calculations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
Simple and systematic random sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Detection of disease . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

3
8 Diagnostic tests 78
8.1 Accuracy and precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Precision . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
8.2 Diagnostic test evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Sensitivity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Specificity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
Positive predictive value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
Negative predictive value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
8.3 Prevalence estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
8.4 Diagnostic strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Parallel interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Serial interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.5 Screening versus diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
8.6 Likelihood ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86

9 Outbreak investigation methods 89

9.1 Verify the outbreak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
What is the illness? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
Is there a true excess of disease? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
9.2 Investigating an outbreak . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Establish a case definition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Enhance surveillance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Describe outbreak according to individual, place and time . . . . . . . . . . . . . . . . . . . . 90
Develop hypotheses about the nature of exposure . . . . . . . . . . . . . . . . . . . . . . . . 91
Conduct analytical studies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
9.3 Implement disease control interventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92

10 Critical appraisal 94
10.1 Description of the evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
10.2 Internal validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Non-causal explanations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Causal explanations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
10.3 External validity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
Can the results be applied to the eligible population? . . . . . . . . . . . . . . . . . . . . . . . 96
Can the results be applied to the source and external populations? . . . . . . . . . . . . . . . 96
10.4 Comparison with other evidence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Consistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Plausibility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Coherency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

4
11 Exercise: outbreak investigation 99
11.1 The problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
11.2 Diagnosis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
11.3 Measures of disease frequency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
11.4 Investigation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
11.5 Measures of association . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
11.6 Recommendations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
11.7 Clinical trial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
11.8 Financial impact . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

12 Review questions 103

12.1 Host, agent, environment . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
12.2 Measures of health . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
12.3 Measures of association . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
12.4 Study design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
12.5 Diagnostic tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105

13 Resources 106

14 Epdemiology formulae sheet 107

14.1 Measures of association . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
14.2 Diagnostic tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
14.3 Sampling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

5
1 Introduction

By the end of this unit you should be able to:

• Compare and contrast clinical and epidemiological approaches to disease management.

• Describe the factors that influence the presence of disease in individuals.
• Describe the factors that influence the presence of disease in populations.
• Explain what is meant by a common source and propagated epidemic (with examples). Explain how you would
distinguish between a common source and propagated epidemic in a disease outbreak situation.

1.1 Definitions and objectives of epidemiology

Epidemiology is the study of diseases in populations. Epidemiologists attempt to characterise those

individuals in a population with high levels of disease and those with low levels. They then ask
questions that help them discover what the high rate group is doing that the low rate group is not
and vice versa. This allows factors influencing the risk of disease to be identified. Once identified,
measures can be applied to reduce exposure to these risk factors — reducing the overall burden
of disease in the population. This allows disease to be controlled even if the precise pathogenic
mechanism (or the aetiologic agent) is not known.
It is useful to distinguish epidemiological from clinical approaches to disease management. The
clinical approach is focussed on individuals and is aimed at diagnosing a disease and then treating
it. It involves physical examination and generation of a list of differential diagnoses. Further ex-
aminations, laboratory tests and possibly response to treatment are then used to narrow the list of
differential diagnoses to a single diagnosis. In an ideal world this will always be the correct diagnosis.
Research in health professionals has shown that the final diagnosis is nearly always drawn from the
initial list of differential diagnoses. If the disease is not on the initial list of differentials then it tends
not to become the final diagnosis. Diseases may be omitted from the list because the clinician is not
familiar with them (exotic or unusual diseases) or because the disease is ‘new’ and has never been
identified before. The epidemiological approach to disease management is conceptually different
in that there is no dependency on being able to precisely define the aetiological agent. It is based
on observing differences and similarities between diseased and non-diseased individuals in order to
try and understand what factors may be increasing or reducing the risk of disease.
In practice, clinicians unwittingly use a combination of clinical and epidemiological approaches in
their day-to-day work. If the problem is relatively clear-cut then an epidemiological approach plays
a very minor role. If the condition is new or more complex then the epidemiological approach is
preferred since it will provide a better understanding of what makes individuals susceptible to disease
and — once these factors are known — the measures required to control the disease become better
defined.

1.2 Host, agent, and environment

Whether or not disease occurs in an individual depends on an interplay of three factors:

• The host;

6
 • The agent; and
• The environment

The host is the individual (animal or human) that may contract a disease. Age, genetic makeup, level
of exposure, and state of health all influence a host’s susceptibility to developing disease. The agent
is the factor that causes the disease (bacteria, virus, parasite, fungus, chemical poison, nutritional
deficiency etc) — one or more agents may be involved. The environment includes surroundings
and conditions either within the host or external to it, that cause or allow disease transmission to
occur. The environment may weaken the host and increase its susceptibility to disease or provide
conditions that favour the survival of the agent.

1.3 Individual, place, and time

The level of disease in a population depends on an interplay of three factors:

• Individual factors: what types of individuals tend to develop disease and who tends to be
spared?

• Spatial factors: where is the disease especially common or rare, and what is different about
these places?

• Temporal factors: how does disease frequency change over time, and what other factors are
associated with these changes?

Individual

Individuals can be grouped or distinguished on a number of characteristics: age, sex, type of hous-
ing, breed, coat colour and so on. An important component of epidemiological research is aimed at
determining the influence of individual characteristics on the risk of disease. Figure 1 shows how
mortality rate for drowning varied among children and young adults in the USA during 1999. The
rate was highest in those aged 1 – 4 years: an age when children are mobile and curious about
everything around them, even though they do not understand the hazards of deep water or how to
survive if they fall in. What conclusions do we draw from this? Mortality as a result of drowning is
highest in children aged 1 – 4 years: preventive measures should be targeted at this age group.

Place

The spatial pattern of disease is typically a consequence of environmental factors. Environmental

factors include aspects of climate (temperature, humidity, rainfall) as well as aspects of management
(e.g. management of animals in a certain area of a country may result in high rates of disease that
may not be seen in other areas). Geographic Information Systems and easy access to spatial data
(e.g. satellite images) have facilitated the ability to conduct spatial epidemiological analyses in recent
years.
Figure 2 shows the geographical distribution of BSE incidence risk in British cattle from July 1992
to June 1993. The amount and type of concentrate feeds fed to cattle is thought to have been
responsible for the higher density of disease in the south of the country, compared with the north.

7
Figure 1: Mortality from drowning by age: USA, 1999. Reproduced from: Hoyert DL, Arias E, Smith BL, Murphy SL,
Kochanek KD (2001) Deaths: final data for 1999. National Vital Statistics Reports volume 49, number 8. Hyattsville MD:
National Center for Health Statistics.

Time

When talking about temporal factors influencing the pattern of disease we need to distinguish be-
tween what we call individual referent time and calendar time. Individual referent time refers to the
timing of events in relation to defined events that occur during a subject’s lifetime. For example,
in dairy cattle medicine we may talk of an increased risk of milk fever during the first 7 days of a
lactation. Here, time is measured in relation to a calving event. Calendar time refers to the absolute
timing of events. We may talk of the number of milk fever cases that occur in August, and compare
those numbers with the number that occur in (say) December.
Temporal patterns of disease in populations are presented graphically using epidemic curves. An
epidemic curve consists of a bar chart showing time on the horizontal axis and the number of new
cases on the vertical axis, as shown in Figure 3. The shape of an epidemic curve can provide
important information about the nature of the disease under investigation. An epidemic occurs when
there is a rapid increase in the level of disease in a population. An epidemic is usually heralded by an
exponential rise in the number of cases in time and a subsequent decline as susceptible animals are
exhausted. Epidemics may arise from the introduction of a novel pathogen (or strain) to a previously
unexposed (naïve) population or as a result of the re-growth of susceptible numbers some time after
a previous epidemic due to the same infectious agent. Epidemics may be described as being either
common source or propagated.
In a common source epidemic, subjects are exposed to a common noxious influence. If the group
is exposed over a relatively short period then disease cases will emerge over one incubation period.
This is called a common point source epidemic. The epidemic of leukaemia cases in Hiroshima
following the atomic bomb blast would be a good example of a common point source epidemic. The
shape of this curve rises rapidly and contains a definite peak at the top, followed by a gradual decline.
Exposure can also occur over a longer period of time, either intermittently or continuously. This
creates either an intermittent common source epidemic or a continuous common source epidemic.
The shape of this curve rises rapidly (associated with the introduction of the agent). The down slope

8
Figure 2: Incidence risk of BSE across Great Britain (expressed as confirmed BSE cases per 100 adult cattle per square
kilometre), July 1992 – June 1993. Reproduced from Stevenson et al. (2000).

of the curve may be very sharp if the common source is removed or gradual if the outbreak is allowed
to exhaust itself.
A propagated epidemic occurs when a case of disease serves as a source of infection for sub-
sequent cases and those subsequent cases, in turn, serve as sources for later cases. In theory,
the epidemic curve of a propagated epidemic has a successive series of peaks reflecting increasing
numbers of cases in each generation. The epidemic usually wanes after a few generations, either
because the number of susceptibles falls below a critical level, or because intervention measures
become effective.
Sometimes epidemic curves can show characteristics of being both common source and propa-
gated. Figure 4 shows the epidemic curve for foot-and-mouth disease in the county of Cumbria
(Great Britain) in 2001. This epidemic started as a common (point) source, then took on the charac-
teristics of a propagative epidemic over time.
Endemic describes the situation when diseases (or events) occur at a predictable frequency. Figure
5 shows data from a descriptive study of dog and cat submissions to a humane shelter in Wellington,
New Zealand from 1999 to 2006. In the plot on the left in Figure 5 there is a marked seasonal
variation in the number of cats submitted to the shelter per month: no such pattern is apparent for
dogs. If data are recorded over extended periods, long-term trends might be evident. In the plot on
the right in Figure 5 it is evident that the number of dogs and cats submitted to the shelter decreased

9
Figure 3: Epidemic curves. The plot on the left is typical of a propagated epidemic. The curve on the right is typical of a
common source epidemic.

steadily throughout the study period.

10
Figure 4: Weekly hazard of foot-and-mouth disease infection for cattle holdings (solid line) and ‘other’ holdings (dashed
line) in Cumbria (Great Britain) in 2001. Reproduced from Wilesmith et al. (2003).

Figure 5: Free-roaming and surrendered dogs and cats submitted to a humane shelter in Wellington, New Zealand, 1999
– 2006 (Rinzin et al. 2008). The plot on the left shows the total number of dogs and cats submitted to the shelter per
calendar month throughout the study period. The plot on the right shows monthly counts of submissions.

11
Figure 6: Descriptive epidemiology of Severe Acute Respiratory Syndrome in Hong Kong, February to April, 2003. A:
Temporal pattern of SARS epidemic in Hong Kong by cluster of infection. B: Spatial distribution of population of Hong
Kong and district-specific incidence (per 10 000 population) over course of epidemic to date. C: Age distribution of
residents of Hong Kong and age-specific incidence (per 10,000 population) over course of epidemic to date. D: Detail of
temporal pattern for Amoy Gardens cluster, according to day of admission, and fitted gamma distribution. Reproduced
from Donnelly et al. (2004).

12
2 Measures of health

By the end of this unit you should be able to:

• Explain (with examples) why it is important to quantify the level of disease in a population.
• Explain the difference between prevalence and incidence, using examples.
• Describe the difference between incidence risk and incidence rate and explain when one measure might be
preferred over the other.
• Describe the difference between closed and open populations, using examples.
• Calculate incidence risk and incidence rate for closed and open populations, given the appropriate data and
formulae.
• Explain why adjusting disease frequency measures is useful in veterinary epidemiology, using examples.

A fundamental task in epidemiological research is to quantify the occurrence of disease. This can
be done by counting the number of affected individuals. To compare levels of disease among groups
of individuals, time frames and locations, we need to consider counts of cases in context of the size
of the population from which those cases arose.
Quantifying the levels of disease in a population is important since it allow health authorities to:

• Determine which diseases are of economic importance.

• Set priorities for the use of resources for disease control activities.

• Plan, implement and evaluate disease control programmes.

• Meet reporting requirements of international organisations (such as the Office International

des Epizooties).

• Demonstrate disease freedom to trading partners.

Before discussing the methods for quantifying disease frequency we define some key terms.
A proportion is a fraction in which the numerator is included in the denominator. Say we have a
herd of 100 cattle and over a 12-month period we identify 58 diseased animals. The proportion of
diseased animals is 58 ÷ 100 = 0.58 = 58%.
A ratio defines the relative size of two quantities expressed by dividing one (numerator) by the other
(denominator). The odds of disease (a ratio) in our herd of 100 cattle is 58:42 or 1.4 to 1.
A closed population is a populaton where no additions or removals occur during a defined follow-
up period. An open population is a population where individuals are added (e.g. as births or
purchases) and removed (e.g. as sales or deaths) during the follow-up period. Most of the popula-
tions you come across in medical and veterinary practice are open.
The term morbidity is used to refer to the extent of disease or disease frequency within a defined
population. Morbidity can be expressed as either prevalence or incidence.

13
2.1 Prevalence

Strictly speaking, the term prevalence refers to the number of cases of a given disease or attribute
that exists in a population at a specified point in time. Prevalence risk is the proportion of a population
that has a specific disease or attribute at a specified point in time. Many authors use the term
‘prevalence’ when they really mean prevalence risk, and these notes will follow this convention.

Number of existing cases

Prevalence = (1)
Size of population

Two types of prevalence are reported in the epidemiological literature: (1) point prevalence equals
the proportion of a population in a diseased state at a single point in time (a snapshot), (2) period
prevalence equals the proportion of a population with a given disease or condition over a specific
period of time. When calculating period prevalence the number of cases equals the number of
individuals which have the disease at the start of the period plus the number of new cases that
occur during the remainder of the follow-up period.
In 1944 the cities of Newburgh and Kingston, New York agreed to participate in a study of the effects of water fluoridation
for prevention of tooth decay in children (Ast and Schlesinger 1956). In 1944 the water in both cities had low fluoride
concentrations. In 1945, Newburgh began adding fluoride to its water — increasing the concentration ten-fold while
Kingston left its supply unchanged. To assess the effect of water fluoridation on dental health, a survey was conducted
among school children in both cities during the 1954 – 1955 school year. One measure of dental decay in children 6
– 9 years of age was whether at least one of a child’s 12 deciduous cuspids or first or second deciduous molars was
missing or had clinical or X-ray evidence of tooth decay.
Of the 216 first-grade children examined in Kingston, 192 had evidence of tooth decay. Of the 184 first-grade children
examined in Newburgh 116 had evidence of tooth decay. Assuming complete survey coverage, there were 192 prevalent
cases of tooth decay among first-grade children in Kingston at the time of the study. The prevalence of tooth decay was
192 ÷ 216 = 89 cases per 100 children in Kingston and 116 ÷ 184 = 63 cases per 100 children in Newburgh.
Reference: Ast DB, Schlesinger ER (1956). The conclusion of a ten-year study of water fluoridation. American Journal
of Public Health, 46: 265-271.

2.2 Incidence

Incidence measures how frequently initially susceptible individuals become disease cases as they
are observed over time. An incident case occurs when an individual changes from being susceptible
to being diseased. The count of incident cases is the number of such events that occur in a defined
population during a specified time period. There are two ways to express incidence: incidence risk
and incidence rate.

Incidence risk

Incidence risk (also known as cumulative incidence) is the proportion of initially susceptible individ-
uals in a population who become new cases during a defined follow-up period.

Number of incident cases

Incidence risk = (2)
Number of individuals initially at risk

14
Incidence risk is reported as the number of cases of disease per head of population over a specified
follow-up period. The follow-up period may be arbitrarily fixed (e.g. the 5-year incidence risk of arthri-
tis) or it may vary among individuals (e.g. the lifetime incidence risk of arthritis). In an investigation
of a localised epidemic the follow-up period may be simply defined as the duration of the epidemic.
Last year a herd of 121 cattle were tested for tuberculosis using the tuberculin test and all tested negative. This year the
same 121 cattle were tested and 25 tested positive.
The incidence risk of tuberculosis in this herd was 21 cases per 100 cattle for the 12-month follow-up period.

Calculating incidence risk for closed populations is straightforward. The denominator is simply the
number of disease free individuals present at the start of the follow-up period.
For open populations things are a little more complicated: we need to take into account those
individuals that enter and leave the population throughout the follow-up period. To do this we take
the number of individuals present at the start, add half of the number that enter the population
during the follow-up period (e.g. births and purchases) and subtract half the number that are lost
(i.e. individuals that leave the population for reasons unrelated to the disease of interest). In effect
this gives the population size at the mid-point of the follow-up period (assuming individuals enter and
exit the population at a constant rate). If an individual can only experience one episode of disease
we include diseased individuals with the group that leave (i.e. once they’ve become a case they are
removed from the population at risk).

• Number at risk = population size at the mid-point of the follow-up period.

• Number at risk = [Nstart + 12 Nnew ] − [ 12 Nlost ].

• Number at risk = [Nstart + 21 Nnew ] − [ 12 (Nlost + Ncases )]. This approach assumes that only
one case of disease is considered per individual.

A pig farm has 125 sows. On 10 March the first case of Actinobacillus pleuropneumoniae is diagnosed on this farm.
Between 10 March and 12 July (124 days) a total of 68 pigs develop clinical signs of Actinobacillus pleuropneumoniae
and are treated. A total of 24 of these 68 pigs require repeat treatments (that is, they recovered from the disease following
first treatment, then got sick again some weeks later and required a second round of treatment). During the outbreak 4
affected pigs die. Calculate the incidence risk of Actinobacillus pleuropneumoniae during the follow up period.
Here we (arbitrarily) make the decision to use the number of affected pigs as the outcome (as opposed to the number
of cases of actinobacillosis). Once a pig becomes a case it is no longer susceptible to disease.
Number at risk = [125 + (0.5 × 0)] − [0.5(0 + 68)]
Number at risk = (125 - 34) pigs
Number at risk = 91 pigs
There were (68 ÷ 91) = 75 affected pigs per 100 sows at risk for the 124 day follow up period.

Incidence rate

Incidence rate (also known as incidence density) is the number of new cases of disease that occur
per unit of individual time at risk during a defined follow-up period.

Number of incident cases

Incidence rate = (3)
Amount of at-risk experience

15
 Table 1: Hypothetical mastitis data.

ID Details Events Days at risk

1 Calve 01 Aug, mastitis 15 Aug, mastitis 15 Sep, mastitis 15 Oct, sold 15 Nov 3 106 a
2 Calve 01 Aug, mastitis 15 Nov, dry off 15 May, 1 365
3 Purchased 01 Dec, mastitis 01 Jan, Dry off 15 May 1 243
4 Calve 01 Aug, Sold 16 Nov 0 107
5 Calve 01 Oct, Died 05 Oct 0 4
Total 5 825

a 01 August to 15 November = 106 days.

Because the denominator is expressed in units of animal- or person-time at risk those individuals
that are withdrawn or are lost to follow-up are easily accounted-for. Consider a study of clinical
mastitis in five cows over a 12-month period, as shown in Table 1.
On the basis of the data presented in Table 1 the incidence rate of clinical mastitis for the 12-month
period is 5 cases per 825 cow-days at risk (equivalent to 2.2 cases of clinical mastitis per cow-year
at risk).
For closed populations the amount of at-risk experience (the denominator) is the number of disease
free individuals present at the start of the follow-up period multiplied by the length of the follow-up
period. For open populations we take the number of individuals present at the start, add half of the
number that enter the population during the follow-up period and subtract half the number that leave
(just as we did when calculating incidence risk for an open population). This number — effectively
the population size at the mid-point of the follow-up period — is then multiplied by the length of the
follow-up period to provide an estimate of the total at-risk experience.

• At-risk experience = population size at the mid-point of the study period × length of study
period.

• At-risk experience = {[Nstart + 21 Nnew ] − [ 12 Nlost ]} × length of study period.

• At-risk experience = {[Nstart + 12 Nnew ] − [ 21 (Nlost + Ncases )]} × length of study period. This
approach assumes that only one case of disease is considered per individual.

Herd management software packages should be able to calculate the exact amount of at-risk expe-
rience because the date of entry and exit is known for each individual member of the population and
it is a simple job to sum the at-risk experience for each individual to yield the total at-risk experience
for the population. The method described here should be used when you want to estimate incidence
rate on the basis of summary data (i.e. when the only information you have is the total number
of animals present at the start of the follow-up period, the total number of additions and the total
number of removals).

16
Gardner et al. (1999) studied on-the-job back sprains and strains among 31,076 material handlers employed by a large
retail merchandising chain. Payroll data for a 21-month period during 1994 – 1995 were linked with job injury claims. A
total of 767 qualifying back injuries occurred during 54,845,247 working hours, yielding an incidence rate of 1.40 back
injuries per 100,000 worker-hours.
Reference: Gardner LI, Landsittel DP, Nelson NA (1999). Risk factors for back injury in 31,076 retail merchandise store
workers. American Journal of Epidemiology, 150: 825 - 833.

A pig farm has 125 sows. On 10 March the first case of Actinobacillus pleuropneumoniae is diagnosed on this farm.
Between 10 March and 12 July (124 days) a total of 68 pigs develop clinical signs of Actinobacillus pleuropneumoniae
and are treated. A total of 24 of these 68 pigs require repeat treatments (that is, they recovered from the disease following
first treatment, then got sick again some weeks later and required a second round of treatment). During the outbreak 4
affected pigs die. Calculate the incidence rate of Actinobacillus pleuropneumoniae during the follow up period.
We make the decision to use the number of actinobacillus cases (as opposed to the number of affected pigs) as the
outcome. The total number of actinobacillus cases was [(68 - 24) + (24 × 2)] = 92. Because we’re using the number of
disease events for the numerator it follows that we account for recovery when calculating the amount of at-risk experience
(the denominator).
At-risk experience = [125 + (0.5 × 0)] − [0.5(0 + 4)] × 124
At-risk experience = (123 × 124) sow-days
At-risk experience = 15,252 sow-days
There were (92 ÷ 15,252) = 0.60 actinobacillosis cases per 100 sow-days at risk for the 124 day follow up period.

Correctly estimating the size of the population at risk (the denominator) presents the most difficulties
when calculating incidence. Remember the following rules of thumb:

• If the population is closed the population at risk equals the number of disease free individuals
present at the start of the follow-up period.

• If the population is open the population at risk should be adjusted to account for those that
enter and leave the population throughout the follow-up period.

The relationship between prevalence and incidence

Table 2 compares the main features of the three measures of disease frequency we have defined.
Figure 7 provides a worked example. This example is based on a herd of 10 animals which are all
disease-free at the beginning of the observation period and followed for a 12-month period. Disease
status is assessed at monthly intervals.
The relationship between point prevalence, period prevalence and incidence can be explained using an analogy with
photography. Point prevalence is like a flashlit photograph: what is happening at an instant in time. Period prevalence
is analogous to a long exposure: the number of events recorded in the photo whilst the camera shutter was open. In a
movie each frame records an instant (point prevalence). By looking from frame to frame one notices new events (incident
events) and can relate the number of such events to a time period (number of frames) to produce incidence rate.

Non-epidemiologists tend to have difficulty in understanding incidence rate so in some situations

it may be useful to convert incidence rate estimates to incidence risk (so you’re talking about the
number of cases of disease per head of population, which is slightly easier to understand). Providing
incidence rate is constant throughout the follow-up period, incidence risk equals:

• Closed population: incidence risk = incidence rate × length of study period.

• Open population (when the study period is short): incidence risk ∼ incidence rate ×
length of study period.

17
 Figure 7: Calculation of prevalence, incidence risk and incidence rate (using exact and approximate methods).

• Open population: incidence risk = 1 − exp(-incidence rate×length of study period)

Prevalence can be estimated from incidence rate, providing incidence rate is constant throughout
the follow-up period and the population is closed:

• Prevalence = (incidence rate × duration of disease) ÷ (incidence rate ×

duration of disease + 1).

• Duration of disease = (prevalence) ÷ (incidence rate × 1 - prevalence).

In a herd of dairy cows the incidence rate of lameness is estimated to be 0.006 cases per cow-day at risk. The average
duration of disease is 7 days.
The estimated prevalence of disease is (0.006 × 7) ÷ (0.006 × 7 + 1) = 0.041
The estimated prevalence of disease is 4.1 cases per 100 cows.

18
 Table 2: A comparison of the main features of prevalence, incidence risk, and incidence rate.

Point prevalence Period prevalence Incidence risk Incidence rate

Numerator All cases counted on a Cases present at period New cases during follow- New cases during follow-
single occasion start + new cases during up period up period
follow-up period
Denominator All individuals examined All individuals examined All susceptible individuals Sum of time period at risk
present at the start of the for susceptible individuals
study present at the start of the
study
Time Single point or period Defined follow-up period Defined follow-up period Measured for each indi-
vidual from beginning of
study until disease event,
exit from the population,
or end of the follow-up
period
Study type Cross-sectional Cohort Cohort Cohort
Interpretation Probability of having dis- Probability of having Probability of developing How quickly new cases
ease at a given point in disease over a defined disease over a defined develop over a defined
time follow-up period follow-up period follow-up period

2.3 Other measures of health

Attack rate

Attack rate is the proportion of a population developing illness during a finite period of time. Attack
rate is equivalent to incidence risk. It is typically used as a measure of average risk during common
source disease outbreaks. Crude attack rate equals the number of cases divided by the number
of individuals potentially exposed. A risk factor specific attack rate equals the number exposed
and develop illness divided by the number exposed. ‘Attack risk’ would be a better term for this
parameter.

Secondary attack rate

Secondary attack rate provides a measure of a disease’s infectiousness. The assumption is that
there is spread of an agent within a group of individuals (e.g. a herd or family) and that not all
cases are a result of a common-source exposure. The numerator for secondary attack rate is the
number of individuals exposed to the primary cases during their infectious period who become ill.
The denominator is the total number of individuals exposed to the primary cases. Again, ‘secondary
attack risk’ would be a better term for this parameter.

Crude mortality rate

Mortality risk (or rate) is an example of incidence where death is the outcome of interest. Cause-
specific mortality risk is the incidence risk of fatal cases of a particular disease in a population at

19
risk of death from that disease. The denominator includes both prevalent cases of the disease
(individuals with disease that haven’t died yet) as well as individuals who are at risk of developing
the disease.

Age-specific mortality rate

Case fatality rate

Case fatality risk (or rate) refers to the incidence of death among individuals who develop the dis-
ease.
Case fatality risk reflects the prognosis of disease among cases, while mortality reflects the burden of deaths from the
disease in the population as a whole.

Proportional mortality rate

As its name implies, proportional mortality is the proportion of all deaths that are due to a particular
cause for a specified population and time period:

Number of deaths from the disease

Proportional mortality = (4)
Number of deaths from all causes

2.4 Adjusted measures of health

Often, we want to compare the frequency of disease in different populations (e.g. herds, regions,
countries). However, because the frequency of disease in a population often depends on other
factors (e.g. age, production type, breed) a higher incidence of disease in one population may
simply reflect the presence of large numbers of individuals where these other factors are present.
Say we’re interested in estimating the prevalence of hypertension in humans. Two areas are selected
for investigation. The first is a regional centre with a large university. The second is a coastal area in
a warmer area of the country which has a large number of retirees. We find that the prevalence of
hypertension varies markedly, with a much higher prevalence in the coastal area. It could be argued
that the reason for the observed difference is the different age distributions of the two populations:
predominantly younger people in the regional centre (students attending the university) and large
numbers of older individuals (retirees) in the coastal area. To make a valid comparison of the two
areas we need to adjust our prevalence estimates to account for the effect of age. Disease frequency
estimates computed using these techniques are referred to as age-adjusted or age-standardised.
There are two methods for adjusting disease frequency estimates: direct adjustment and indirect
adjustment.

20
Direct adjustment

With direct adjustment the adjusted count for the ith strata equals the observed disease frequency
estimate (i.e. prevalence or incidence) multiplied by a standard population estimate for the ith strata:

Directly adjusted counti = OBS Ri × STD Pi (5)

Where:
OBS Ri : the observed prevalence or incidence in the ith strata
STD Pi : the size of the standard population in the ith strata
Another example. Say we’re interested in comparing the incidence of Johne’s disease in two regions.
Counts of Johne’s cases recorded over a 12-month period and counts of animals, stratified by age,
are shown in Table 3.
Table 3: Details of Johne’s disease events recorded over a 12-month period in two regions.

Age group Region A Region B

n Johne’s n in group IR a n Johne’s n in group IR a
Young 1 1000 10 20 10,000 20
Middle aged 25 5,000 50 50 5,000 100
Old 100 10,000 100 20 1000 200
Total 126 16,000 79 90 16,000 56

a Incidence risk (cases per 10,000).

The incidence risk of Johne’s disease, by age, is much higher in Region B than Region A. Inter-
estingly, when we look at the population totals, the incidence of disease is greater in Region A (79
cases per 10,000) compared with Region B (56 cases per 10,000). Figure 8 helps to explain why
this is the case. In Figure 8 the open circles show, for each region, the incidence risk of Johne’s
disease as a function of the three categories of age. Superimposed over each open circle is a box
with size proportional to the size of the population in each region-age group category. Clearly seen
here is that in Region A there are relatively large numbers of older animals whereas in Region B
there are relatively large numbers of younger animals. The summary estimate of Johne’s incidence
risk for each region is indicated on each line as a filled circle: 79 cases per 100 in Region A and 56
cases per 100 in Region B. Region A has a high summary estimate of Johne’s disease incidence
because the summary figure has been weighted (‘dragged upwards’) by the large numbers of older
animals that predominate in Region A’s population. Similarly, Region B has a low summary estimate
of Johne’s disease incidence because the summary figure has been weighted by the large number
of younger animals present.
What we say in this situation is that age has confounded the true association between region and
Johne’s disease incidence.

21
Figure 8: Incidence risk of Johne’s disease recorded over a 12-month period in two regions. The open circles show
the incidence risk for each age group. The boxes indicate the size of the population contributing to each incidence risk
estimate. The solid circles show the regional summary estimates of Johne’s incidence risk.

The Johne’s disease example cited above is a classic case of Simpson’s paradox. Simpson’s paradox refers to the
situation where, due to the presence of a confounding variable, an association is in one direction at the whole group
level, but the opposite direction in each of the subgroups.
Simpson’s paradox is not really a paradox but rather an extreme manifestation of the fact that associations may change
in the presence of confounders.

The confounding effect of age can be removed by directly adjusting the Johne’s incidence risk es-
timates. Direct adjustment involves a three-step process. First, pick a reference population. There
are many options for doing this (e.g. using census data) but one simple approach is to combine the
population counts provided in Table 3, as shown in Table 4.

Table 4: A standard age distribution for the two regions under investigation.

Age group n in group

Young 1000 + 10,000 = 11,000
Middle aged 5000 + 5000 = 10,000
Old 10,000 + 1000 = 11,000
Total 16,000 + 16,000 = 32,000

The second step involves applying the age-specific incidence risk estimates for each region to the
reference population. This provides the directly adjusted Johne’s counts for each strata, as shown
in Table 5.

22
 Table 5: Directly adjusted Johne’s counts.

Age group Region A Region B

Directly adjusted count Directly adjusted count
Young (10 ÷ 10,000) × 11,000 = 11 (20 ÷ 10,000) × 11,000 = 22
Middle aged (50 ÷ 10,000) × 10,000 = 50 (100 ÷ 10,000) × 10,000 = 100
Old (100 ÷ 10,000) × 11,000 = 110 (200 ÷ 10,000) × 11,000 = 220
Total 11 + 50 + 110 = 171 22 + 100 + 220 = 342

The third (and final) step is to divide the total directly adjusted Johne’s counts in each region by the
size of the reference population. This gives 171 ÷ 32,000 = 53 cases per 10,000 for Region A and
342 ÷ 32,000 = 107 cases per 10,000 for Region B. The adjusted incidence risk of Johne’s in Region
B is now greater than that in Region A, consistent with the strata-level risk estimates. The directly
adjusted estimates allow the contribution of region on Johne’s disease incidence to be assessed
independent of the age structure of the regions under investigation.

Indirect adjustment

With indirect adjustment the adjusted count for the ith strata equals a standardised frequency esti-
mate multiplied by the observed population size for the ith strata:

Indirectly adjusted counti = STD Ri × OBS Pi (6)

Where:
STD Ri : the standard incidence or prevalence in the ith strata of the population
OBS Pi : the observed population size in the ith strata
It is usual to set the standardardised incidence (or prevalence) for the ith strata as the sum of the total
number of disease events across all strata divided by the total population size. Using this approach
the indirectly adjusted disease count for the ith strata equals the standardardised incidence (or
prevalence) multiplied by the population size for each strata. What this provides is the expected
number of disease events within each strata (Ei ) assuming the strata-level risk of disease is the
same as the entire population. It is common to divide the observed number of disease events (Oi )
per strata by the expected number (Ei ) to yield a standardised morbidity or mortality ratio (SMR). To
continue the example introduced earlier in Table 6 we present the incidence risk of Johne’s disease
for each region and for both regions combined.
In Table 7 we calculate the expected number of cases (the indirectly adjusted disease count) and
the standardised morbidity ratio for Johne’s disease.
In Table 7 we apply our best estimate of the ‘average’ incidence of Johne’s disease in the population
(i.e. 216 cases in 32,000 animals at risk) to each herd to produce the number of Johne’s cases
expected. The number of cases actually observed is then divided by this expected count to produce
a standardised morbidity ratio. In Region A there was 1.17 times the number of Johne’s cases
expected. In Region B there were 0.84 times the number of Johne’s cases expected. Note that this

23
 Table 6: Counts of cases, size of the population at risk and incidence risk of Johne’s disease in two regions.

Region n Johne’s n in group IR a

A 126 16,000 79
B 90 16,000 56
Total 216 32,000 67

a Incidence risk (cases per 10,000).

Table 7: Observed and expected number of Johne’s cases and the standardised morbidity ratio (SMR) for Johne’s for
each region.

Region Observed Expected SMR

A 126 0.0067 × 16,000 = 107 126 ÷ 107 = 1.17
B 90 0.0067 × 16,000 = 107 90 ÷ 107 = 0.84

approach hasn’t adjusted for the effect of age, so Region B’s SMR is (incorrectly) less than that of
Region A.
If area units (e.g. states, counties, census tracts) are the basis for stratification it is common to
plot the SMR for each area unit in the form of a choropleth map (a map where areas are coloured
according to the value of the outcome of interest). Choropleth maps of SMR estimates are an
effective way to describe the geographical distribution of disease in a population, and how this might
change over time. Figure 9 provides an example of this approach applied to bovine spongiform
encephalopathy in Great Britain.
We know that the prevalence of a given disease throughout a country is 0.01%. If we are presented with a region with
20,000 animals the expected number of cases of disease in this region will be 0.0001 × 20,000 = 2.
If the actual number of cases of disease in this region is 5, then the standardised mortality (morbidity) ratio is 5 ÷ 2 =
2.5. That is, there were 2.5 times more cases of disease in this region, compared with the number of cases expected.

24
 0 - 0.5 0 - 0.5
0.5 - 1.0 0.5 - 1.0
1.0 - 2.0 1.0 - 2.0
2.0 - 6.0 2.0 - 6.0

(a) SMR: pre-control cohort (b) SMR: post-control cohort

Figure 9: An example of the use of indirect standardisation used to describe the change in spatial distribution of disease
risk over time. Choropleth maps of area-level standardised mortality ratios (SMRs) for bovine spongiform encephalopathy
in British cattle 1986 – 1997, for (a) cattle born before the 18 July 1988 ban on feeding meat and bone meal to ruminants,
and (b) cattle born between 18 July 1988 and 30 June 1997. The above maps show a shift in area-level risk over time
towards the east of the country (even though the incidence of BSE reduced markedly from 1988 to 1997). Reproduced
from Stevenson et al. (2005).

25
3 Study design

By the end of this unit you should be able to:

• Describe the difference between descriptive and analytical epidemiological studies (giving examples of each).
• Describe the major features of the following study designs: case reports, case series, descriptive studies, eco-
logical studies, cross-sectional studies, cohort studies, case-control studies, clinical trials, randomised clinical
trials, and community trials.
• Suggest an appropriate study design to identify risk factors for disease, given details of a disease problem in a
population of animals. Be able to justify your chosen design.
• Describe the strengths and weaknesses of cross-sectional studies, cohort studies, case-control studies, and
clinical trials.

A study generally begins with a research question. Once the research question has been specified
the next step is to choose a study design. A study design is a plan for selecting study subjects and
for obtaining data about them. Figure 10 shows the major types of epidemiological study designs.
There are three main study types: (1) descriptive studies, (2) analytical studies, and (3) experimental
studies.

Figure 10: Tree diagram outlining relationships between the major types of epidemiologic study designs.

Descriptive studies are those undertaken without a specific hypothesis. They are often the earliest
studies done on a new disease in order to characterise it, quantify its frequency, and determine how
it varies in relation to individual, place and time. Analytical studies are undertaken to identify and test
hypotheses about the association between an exposure of interest and a particular outcome. Exper-
imental studies are also designed to test hypotheses between specific exposures and outcomes —
the major difference is that in experimental studies the investigator has direct control over the study
conditions.

3.1 Descriptive studies

The hallmark of a descriptive study is that it is undertaken without a specific hypothesis.

26
Case reports

A case report describes some ‘newsworthy’ clinical occurrence, such as an unusual combination of
clinical signs, experience with a novel treatment, or a sequence of events that may suggest previ-
ously unsuspected causal relationships. Case reports are generally reported as a clinical narrative.
Trivier at al (2001) reported the occurrence of fatal aplastic anaemia in an 88 year-old man who had taken clopidogrel,
a relatively new drug on the market that inhibits platelet aggregation. The authors speculated that his fatal illness may
have been caused by clopidogrel and wished to alert other clinicians to a possible adverse effect of the drug.
Reference: Trivier JM, Caron J, Mahieu M, Cambier N, Rose C (2001). Fatal aplastic anaemia associated with clopido-
grel. Lancet, 357: 446.

Cases series

Whereas a case report shows that something can happen once, a case series shows that it can
happen repeatedly. A case series identifies common features among multiple cases and describes
patterns of variability among them.
After bovine spongiform encephalopathy (BSE) appeared in British cattle in 1987, there was concern that the disease
might spread to humans. A special surveillance unit was set up to study Creutzfeld-Jacob disease (CJD), a rare and fatal
progressive dementia that shares clinical and pathological features of BSE. In 1996 investigators at the unit described
ten cases that met the criteria for CJD but had all occurred at unusually young ages, showed distinctive symptoms and,
on pathological examination, had extensive prion protein plaques throughout the brain similar to BSE.
Reference: Will RG, Ironside JW, Zeidler M, Cousens SN, Estibeiro K, Alperovitch A (1996). A new variant of Creutzfeld-
Jacob disease in the UK. Lancet, 347: 921 - 925.

Descriptive studies based on rates

Descriptive studies based on rates quantify the burden of disease on a population using incidence,
prevalence, mortality or other measures of disease frequency. Most use data from existing sources
(such as birth and death certificates, disease registries or surveillance systems). Descriptive studies
can be a rich source of hypotheses that lead later to analytic studies.
Schwarz et al. (1994) conducted a descriptive epidemiological study of injuries in a predominantly African-American
part of Philadelphia. An injury surveillance system was set up in a hospital emergency centre. Denominator information
came from US census data. These authors found a high incidence of intentional interpersonal injury in this area of the
city.
Reference: Schwarz DF, Grisso JA, Miles CG, Holmes JH, Wishner AR, Sutton RL (1994). A longitudinal study of injury
morbidity in an African-American population. Journal of the American Medical Association, 271: 755 - 760.

3.2 Analytical studies

Analytical studies are undertaken to test a hypothesis. In epidemiology the hypothesis typically
concerns whether a certain exposure causes (or is assoicated with) a certain outcome — e.g. does
cigarette smoking cause lung cancer? The term exposure is used to refer to any trait, behaviour,
environmental factor or other characteristic as a possible cause of disease. Synonyms for exposure
are: potential risk factor, putative cause, independent variable, and predictor. The term outcome

27
generally refers to the occurrence of disease. Synonyms for outcome are: effect, end-point, and
dependent variable.
The hypothesis in an analytic study is whether an exposure actually causes an outcome (not merely
whether the two are associated). Each of Hill’s criteria for causation are usually required to be met to
support a case for causality, but probably the most important of these is that exposure must precede
the outcome in time. See Chapter 6
for details.

Ecological studies

In an ecological study the unit of analysis is a group of individuals (such as counties, states, cities, or
census tracts). Summary measures of exposure and summary measures of outcome are compared
and inference is made at the individual level.
Yang et al. (1998) conducted an ecological study examining the association between chlorinated drinking water and
cancer mortality among 28 municipalities in Taiwan. The investigators found a positive association between the use of
chlorinated drinking water and mortality from rectal, lung, bladder, and kidney cancer.
Reference: Yang CY, Chiu HF, Cheng MF, Tsai SS (1998). Chlorination of drinking water and cancer in Taiwan. Environ-
mental Research, 78: 1 - 6.

Ecological studies are relatively quick and inexpensive to perform and can provide clues to possible
associations between exposures and outcomes of interest. A major disadvantage of ecological
studies is that of ecological fallacy: the assumption that an observed relationship in aggregated data
will hold at the individual level.

Cross-sectional studies

In a cross-sectional study a sample of individuals from a population is taken at a single or relatively

short period of time. Individuals included in the sample are examined for the presence of disease and
their status with regard to the presence or absence of specified exposures. Cross sectional studies
commonly involve surveys to collect data. Surveys range from simple one-page questionnaires
addressing a single variable, to highly complex, multiple page designs.

Figure 11: Schematic diagram of a cross-sectional study.

28
Advantages: Cross-sectional studies can be carried out quickly and tend to be substantially cheaper
than other study designs (because there is no requirement to follow study subjects for extended
periods of time).
Disadvantages: Cross-sectional studies cannot provide information on the incidence of disease in
a population — only an estimate of prevalence. Not suited for diseases of short duration — imagine
doing a cross-sectional study to determine the prevalence of milk fever in dairy cows. Difficult to
investigate or establish cause and effect relationships because it can be difficult to determine if the
exposure occurred before the outcome of interest.
Scuffham et al. (2008) conducted a cross-sectional study of 867 New Zealand veterinarians to identify risk factors for
work-related musculoskeletal discomfort. Eighteen percent of those who completed an on-line survey reported a period
of absence from work during the previous 12 months due to musculoskeletal discomfort. The odds of musculoskeletal
discomfort was 2.64 times (95% CI 1.08 – 6.30) greater in veterinarians who reported a high level of disatisfaction with
the difficulty of their working conditions.
Which came first? Did disatisfaction with working conditions increase the likelihood of developing musculoskeletal
discomfort, or did the presence of musculoskeletal discomfort then produce disatisfation with working conditions? This
type of uncertainty around cause and effect relationships is an example of one of the disadvantages of a cross-sectional
study design.
Reference: Scuffham AM, Legg SJ, Firth EC, Stevenson MA (2009). Prevalence and risk factors for musculoskeletal
discomfort in New Zealand veterinarians. Applied Ergonomics, 41: 444 - 453.

Cohort studies

A cohort study involves comparing disease incidence over time between groups (cohorts) that are
found to differ on their exposure to a factor of interest. Cohort studies are either prospective or
retrospective (Figure 12).

Figure 12: Schematic diagram of a prospective and retrospective cohort study.

A prospective cohort study begins with the selection of two groups of non-diseased animals, one
exposed to a factor postulated to cause a disease and the other unexposed. The groups are followed
over time and their change in disease status is recorded during the study period. A retrospective

29
cohort study starts when all of the disease cases have been identified. The history of each study
participant is carefully evaluated for evidence of exposure to the agent under investigation.
Advantages: Because subjects are monitored over time for disease occurrence, cohort studies
provide estimates of the absolute incidence of disease in exposed and non-exposed individuals.
By design, exposure status is recorded before disease has been identified. In most cases, this
provides unambiguous information about whether exposure preceded disease. Cohort studies are
well-suited for studying rare exposures. This is because the relative number of exposed and non-
exposed persons in the study need not necessarily reflect true exposure prevalence in the population
at large.
Disadvantages: Prospective cohort studies require a long follow-up period. In the case of rare
diseases large groups are necessary. Losses to follow-up can become an important problem. Often
quite expensive to run.
To assess the possible carcinogenic effects of radio-frequency signals emitted by cellular telephones, Johansen et al.
(2001) conducted a retrospective cohort study in Denmark. Two companies that operate cellular telephone networks
provided names and addresses for all 522,914 of their clients for the period 1982 to 1995. The investigators matched
these records to the Danish Central Population Register. After cleaning the data 420,095 cellular telephone subscribers
remained and formed the exposed cohort. All other Danish citizens during the study years became the unexposed
cohort. The list of exposed and unexposed individuals were then matched with the national cancer registry. The resulting
data allowed calculation of cancer incidence rates. Overall, 3,391 cancers had occurred among cellular telephone
subscribers, compared with 3,825 cases expected based on age, gender, and calendar-year distribution of their person
time at risk.
Reference: Johansen C, Boise J, McLaughlin J, Olsen J (2001). Cellular telephones and cancer — a nationwide cohort
study in Denmark. Journal of the National Cancer Institute, 93: 203 - 237.

Case-control studies

Say we’re interested in investigating risk factors for a rare disease such as bladder cancer in dogs.
Imagine we have access to a perfect data set where we have the medical records for every dog in
the country and details about things these dogs have been exposed to in their first year of life. For a
given exposure (e.g. access to benzidine) we can present the data in a 2 × 2 table format, as shown
in Table 8.
Table 8: Hypothetical data from a study of bladder cancer in a population of dogs.

Disease + Disease - Total

Benzidene + 60 188,940 189,000
Benzidene - 57 278,943 279,000
Total 117 467,883 468,000

In this hypothetical example the risk of bladder cancer is 32 per 100,000 in those dogs exposed to
benzidene in their first year of life, compared with 20 per 100,000 in those not exposed. The risk of
bladder cancer is 1.6 times greater in exposed dogs than non-exposed dogs.
Now think of the logistics involved in carrying out this study. We would have to enroll 468,000 dogs,
ask detailed questions about their management during the first year of life, then follow them for an

30
extended period (years) to work out which of them got bladder cancer. A formidable task. A case-
control design is intended to provide the same answer in a much simpler way by studying all of the
dogs who got bladder cancer and a sample of dogs who did not.

Table 9: Hypothetical data from a study of bladder cancer in a population of dogs. The data is comprised of 117 cases of
bladder cancer and 468 controls selected at random from the population.

Disease + Disease - Total

Benzidene + 60 189 249
Benzidene - 57 279 336
Total 117 468 585

Suppose we used a case-control approach where we investigate all 117 dogs with bladder cancer
and a sample of controls chosen by selecting one out of every 1000 dogs who remained free of
disease. The results are shown in Table 9. If we only had access to the information provided in
Table 9 we wouldn’t be able to calculate the risk of bladder cancer in either exposed or unexposed
dogs because we don’t know the size of the population at risk. What we can do however is work out
the odds of cancer in the exposed and unexposed groups and compare them. The odds of cancer
in benzidene exposed dogs is 60 ÷ 189 = 0.32. The odds of cancer in non-exposed dogs is 57 ÷
279 = 0.20. The ratio of theses two odds is 0.32 ÷ 0.20 = 1.6, which is the same as the risk ratio
calculated earlier. The reason we got the same result should be obvious: 60 ÷ 189 divided by 57 ÷
279 is the same as 60 ÷ 189,000 divided by 57 ÷ 279,000. It is the same ratio as before — the two
denominators have simply been divided by 1000 as we have sampled only one out of every 1000
dogs who did not get bladder cancer.
This example demonstrates the usefulness of the case-control study design. Details collected on
all identified cases and a selection of disease-negative animals (‘controls’) yields the same result
as a very expensive (and usually impractical) study where every member of a population at risk is
examined.
In a case-control study a group of cases and non-cases (‘controls’) are selected and we compare the
frequency of exposure factors in the cases with that of the controls. Cases are those study subjects
who have developed the outcome of interest whereas controls are those who have not developed
the outcome of interest at the time of selection. The key thing is that the set of controls represent
a set of individuals whose exposure to the factor of interest reflects the exposure in the population
from which the cases were drawn. In most situations the individual is the unit of interest, this design
applies equally as well to aggregates of individuals (such as litters, pens, and herds). Figure 13 is a
diagramatic representation of the case-control design.
The key issue when designing a case-control study is to ensure that cases and controls are simi-
lar in every way except for the exposure factors hypothesised to be associated with the disease of
interest. Controls should be drawn from the same general population as the cases — this is neces-
sary to protect against the possible distortions from effect modifiers (confounders). There are three
approaches that might be used to ensure that cases and controls are similar:

• Restricted sampling. If breed is a likely confounder you might only select one breed in the
study (the dominant breed in the source population).

31
 Figure 13: Schematic diagram of a case-control study.

• Matching. Each case is matched with a control that has identical (or at least similar) values
of the confounding variable (e.g. age and sex). This method provides direct control over
known confounders and under certain conditions the efficiency of the analysis is improved.
Disadvantages: (1) recruitment of suitable controls can be difficult (when it is difficult to find
a suitable match); (2) you cannot quantify the effect of the matching variable on the risk of
disease; (3) analysis of the data must take into account the effect of matching; (4) it is possible
to overmatch, which decreases the efficiency of the study (and sometimes introduces bias).

• Analytical control. Multivariable regression techniques may be applied to remove the effect of
known confounders.

Advantages: Case-control studies are an efficient method for studying rare diseases. Because
subjects have experienced the outcome of interest at the start of the study, case-control studies are
quick to run and are considerably cheaper than other study types.
Disadvantages: Case-control studies cannot provide information on the disease incidence in the
studied population. The study is reliant on the quality of past records or recollection of study partic-
ipants. It can also be very difficult to ensure an unbiased selection of the control group and, as a
result, the representativeness of the sample selection process is difficult to guarantee.
Muscat et al. (2000) sought to test the hypothesis that cellular telephone use affects the risk of brain cancer. From
1994 to 1998 at five academic medical centres in the USA they recruited 469 cases aged 18 to 80 years with newly
diagnosed cancer originating in the brain. Controls (n = 422) were inpatients without brain cancer at those hospitals,
excluding those with leukaemia or lymphoma. Controls were sampled to match the cases on age, sex, race and month
of admission. Each case and control was then interviewed about any past subscription to a cellular telephone service.
Overall 14.1% of cases and 18.0% of controls reported ever having had a subscription for a cellular telephone service.
After adjusting for age, sex, race, education, study centre, and month and year of interview, the risk of developing brain
cancer in a cellular telephone user was estimated to be 0.85 (95% CI 0.6 – 1.2) times as great as in a non-user.
Reference: Muscat JE, Malkin MG, Thompson S, Shore RE, Stellman SD, McRee D (2000). Handheld cellular telephone
use and risk of brain cancer. Journal of the American Medical Association, 284: 3001 - 3007.

32
In a cohort study definition of exposure status (exposure-positive, exposure-negative) comes first. Subjects are then
followed over time to determine their outcome status (disease-positive, disease-negative).
In a case-control study outcome status (disease-positive, disease-negative) is defined first. The history provided by
each subject provides information about exposure status (exposure-positive, exposure-negative).

Hybrid study designs

A nested case-control study is similar to a cohort study with the key difference that a sample of
non-cases are selected for analysis (rather than the entire cohort, as in the case of a cohort study).
Figure 14 shows a diagram of a nested case-control design.

Figure 14: Schematic diagram of a nested case-control study.

Advantages: Nested case-control studies are useful when it is either too costly or not feasible
to perform additional analyses on an entire cohort (e.g. if collection of specimens and laboratory
analysis of specimens is expensive). Compared with standard case-control studies, nested studies:
(1) can utilise exposure and confounder data originally collected before the onset of the disease,
thus reducing potential recall bias and temporal ambiguity, and (2) include cases and controls drawn
from the same cohort, decreasing the likelihood of selection bias. The nested case-control study is
thus considered a strong observational study, comparable to its parent cohort study.
Disadvantages: A concern, usually minor, is that the remaining non-diseased persons from whom
the controls are selected when it is decided to do the nested study, may not be fully representative
of the original cohort due to death or losses to follow-up.
To determine if Helicobacter pylori infection was associated with the development of gastric cancer, Parsonnet et al.
(1991) identified a cohort of 128,992 persons who had been followed since the mid-1960s. Of the original cohort, 189
patients developed gastric cancer. The investigators carried out a nested case-control study by selecting all of the 189
gastric cancer patients as cases and another 189 cancer-free individuals from the same cohort as controls. H. pylori
infection status was determined using serum obtained at the beginning of the follow-up period. All total of 84% of the
confirmed gastric cancer cases had been infected previously with H. pylori, while only 61% of the controls had been
infected. This indicated a positive association between H. pylori infection and gastric cancer risk.
Reference: Parsonnet J, Friedman GD, Vandersteen DP, Chang Y, Vogelman JH, Orentreich N, Sibley RK (1991).
Helicobacter pylori infection and the risk of gastric-carcinoma. New England Journal of Medicine, 325(16): 1127 - 1131.

In a case-crossover study a set of cases (subjects) is identified and a period of time before the
onset of disease is selected (termed the case window) wherein the exposure to the risk factor of

33
interest is evaluated. For each subject a second, non-overlapping time window (the control window)
of the same length as the case window is selected, during which the subject did not experience the
disease. This design is suitable for studying the transient effects of exposures that can vary over
time and precipitate acute events (e.g. epilepsy episodes, asthma attacks). The design is efficient
in that each case acts as its own control.
A study was conducted to determine if sleep disturbance was a risk factor for injury in children (Valent et al. 2001). A set
of cases were identified and each child was asked if their sleep was disturbed in the 24 hours before the injury occurred
(the case window) and in the 24 hours before that (the control window). Among 181 boys, 40 had less than 10 hours
sleep on both days; 111 had less than 10 hours on neither day; 21 had less than 10 hours only on the day before the
injury; and 9 had less than 10 hours sleep on the penultimate day before the injury. The odds ratio for injury, comparing
days without and with 10 hours or more sleep was 2.33 (95% CI 1.02 – 5.79).
Reference: Valent F, Brusaferro S, Barbone F (2001). A case-crossover study of sleep and childhood injury. Pediatrics,
107: E23.

A panel study combines the features of cross-sectional and a prospective cohort designs. It can
be viewed as a series of cross-sectional studies conducted on the same subjects (the panel) at
successive time intervals (sometimes referred to as waves). This design allows investigators to
relate changes in one variable to changes in other variables over time. The difference between a
panel study and a prospective cohort study is subtle. In a cohort study information is collected on
subject events as they occur during a follow up period. In a panel study information is collected at
defined periods of time (by interview, for example).
A repeated survey is a series of cross-sectional studies performed over time on the same study
population, but each is sampled independently. Whereas panel studies follow the same individuals
from survey to survey, repeated surveys follow the same study population (which may differ in com-
position from one survey to the next). Repeated surveys are useful for identifying overall trends in
health status over time.

Randomised clinical trials

The randomised clinical trial is the epidemiologic design that most closely resembles a laboratory
experiment. The major objective is to test the possible effect of a therapeutic or preventive interven-
tion. The design’s key feature is that a formal chance mechanism is used to assign participants to
either the treatment or control group. Subjects are then followed over time to measure one or more
outcomes, such as the occurrence of disease. All things being equal, results from randomised trials
offer a more solid basis for inference of cause and effect than results obtained from any other study
design.
Advantages: Randomisation generally provides excellent control over confounding, even by factors
that may be hard to measure or that may be unknown to the investigator.
Disadvantages: For many exposures it may not be ethical or feasible to conduct a clinical trial (e.g.
exposure to pollution). Expensive. Impractical if long periods of follow-up required.

34
 Figure 15: Schematic diagram of a randomised clinical trial.

Bacterial vaginosis affects an estimated 800,000 pregnant women each year in the USA and has been found to be
associated with premature birth and other pregnancy complications. To determine whether treatment with antibiotics
could reduce the incidence of adverse pregnancy outcomes, Carey et al. (2000) screened 29,625 pregnant women to
identify 1953 who had bacterial vaginosis, met certain other eligibility criteria, and consented to participate. Women were
randomly assigned to receive either: (1) two 2 gram doses of metronidazole, or (2) two doses of a similar-appearing
placebo.
Bacterial vaginosis resolved in 78% of women in the treatment group, but in only 37% of women in the placebo group.
Pre-term labour, postpartum infections in the mother or infant, and admission to the neonatal intensive care unit were
equally common in both groups.
Reference: Carey JC, Klebanoff MA, Hauth JC, Hillier SL, Thom EA, Ernest JM et al. (2000). Metronidazole to prevent
preterm delivery in pregnant women with asymptomatic bacterial vaginosis. New England Journal of Medicine, 342: 534
- 540.

Community trials

Instead of randomly assigning individuals to treatment or control groups, community trials assign
interventions to entire groups of individuals. In the simplest situation one group (community) receives
the treatment and another serves as a control.

3.3 Comparison of the major epidemiological study designs

Cohort studies involve enumeration of the denominator of the disease measure (individual time
at risk) while case-control studies only sample from the denominator. Cohort studies therefore
provide an estimate of incidence and risk whereas case-control studies can only estimate ratios.
Prospective cohort studies provide the best evidence for the presence of cause-effect relationships,
because any putative cause has to be present before disease occurs. Since these study designs
are based on observation within a largely uncontrolled environment it is possible that there are still
other unmeasured factors which produce cause-effect relationships that might be identified. The

35
prospective cohort study is inefficient for studying rare diseases, which is a particular strength of the
case-control study. A carefully designed cross-sectional study is more likely to be representative of
the population than a case-control study.

Table 10: Comparison of the features of the cohort, case-control and cross-sectional study designs.

Criteria Cohort Case-control Cross-sectional

Sampling Separate samples of exposed Separate samples of diseased Random sample of study
and non-exposed individuals and non-diseased individuals group
Time Usually prospective (but may Usually retrospective Single point
be retrospective)
Causality Causality through evidence of Preliminary causal hypothesis Association between disease
temporality and risk factor
Risk Incidence risk, incidence rate None Prevalence
Measure of assocation Risk ratio, odds ratio Odds ratio Risk ratio, odds ratio

36
4 Measures of association

By the end of this unit you should be able to:

• Given disease count data, construct a 2 × 2 table and, given the appropriate formulae, explain how to calculate
the following measures of association: risk ratio, odds ratio, attributable risk, attributable fraction, population
attributable risk, population attributable fraction.
• Interpret the following measures of association: risk ratio, odds ratio, attributable risk, attributable fraction, popu-
lation attributable risk, population attributable fraction.
• Describe situations where the risk ratio is not a valid measure of association between an exposure and an
outcome.

An important task in epidemiology is to quantify the strength of association between exposures and
outcomes. In this context the term ‘exposure’ is taken to mean a variable whose causal effect is to
be estimated. Exposures may be harmful, beneficial or both harmful and beneficial (e.g. if an immu-
nisable disease is circulating, exposure to immunising agents helps most recipients but may harm
those who experience adverse reactions). Outcomes are all the possible results that may stem from
exposure to a causal factor or from preventive or therapeutic interventions (Porta, Greenland and
Last 2008). A risk factor is an attribute or exposure that is associated with an increased probability
of a specified outcome, such as the occurrence of a disease. Consider the following:

• Motor vehicles involved in accidents often have worn tyres.

• In humans with coronary heart disease high blood pressure is a common clinical finding.

• The likelihood of clostridial disease is reduced in those animals that have been vaccinated.

In the above examples worn tyres, high blood pressure and vaccination are the exposures. Motor
vehicle accidents, coronary heart disease and clostridial disease are the outcomes. Each of the
exposures are referred to as risk factors (vaccination being a protective risk factor). If we can iden-
tify risk factors for disease then we’re in a good position to make recommendations about health
management. Much of epidemiological research is concerned with identifying and quantifying the
effect of risk factors on the likelihood of disease.
Consider a study where subjects are disease free at the start of the study and all are monitored for
disease occurrence for a specified time period. If both exposure and outcome are binary variables
(yes or no) we can present the counts of subjects in each of the four exposure-disease categories
in a 2 × 2 table.
Table 11: The 2 × 2 table. Disease status (positive and negative) are listed as columns, exposure status (positive and
negative) as rows. Counts of individuals in each of the disease-exposure categories are entered into each of the cells
representing each of the four exposure-disease combinations.

Diseased Non-diseased Total

Exposed a b a+b
Non-exposed c d c+d
Total a+c b+d a+b+c+d

37
To illustrate the use of 2 × 2 tables in practice Table 12 presents the results of a cohort study
investigating the relationship between feeding cats dry food (DCF) and feline lower urinary tract
disease (FLUTD). In this example the feeding of dry cat food is the exposure and the presence of
feline lower urinary tract disease is the outcome.

Table 12: Results from a cohort study investigating the association between dry cat food use and the presence of FLUTD
in cats.

FLUTD + FLUTD - Total

DCF + 13 2163 2176
DCF - 5 3349 3354
Total 18 5512 5530

4.1 Measures of strength

Incidence risk ratio

The incidence risk ratio is defined as the incidence risk of disease in the exposed group divided by
the incidence risk of disease in the unexposed group:
Incidence risk of disease in the exposed population: RE+ = a/(a + b)
Incidence risk of disease in the non-exposed population: RE− = c/(c + d)

RE+
RR = (7)
RE−

The incidence risk ratio provides an estimate of how many times more likely exposed individuals are
to experience disease, compared with non-exposed individuals. If the incidence risk ratio equals 1,
then the risk of disease in the exposed and non-exposed groups are equal. If the incidence risk
ratio is greater than 1, then exposure increases the risk of disease with greater departures from 1
indicative of a stronger effect. If the incidence risk ratio is less than 1, exposure reduces the risk of
disease and exposure is said to be protective. The incidence risk ratio can’t be estimated in case-
control studies, as these studies do not allow calculation of risks. Odds ratios are used instead —
see below.

Incidence rate ratio

In a study where incidence rate has been measured (rather than incidence risk) the incidence rate
ratio can be calculated. This is the ratio of the incidence rate in the exposed group to that in the
non-exposed group. The incidence rate ratio is interpreted in the same way as the incidence risk
ratio.
The term relative risk is used as a synonym for both incidence risk ratio and incidence rate ratio.
Incidence risk ratios and incidence rate ratios range between 0 and infinity.

38
Figure 16: Explanation of risk ratio as it relates to the FLUTD cohort study. In the FLUTD cohort study the incidence risk
of FLUTD in the DCF+ group was 13 ÷ 2176 = 5.97 cases per 1000. The incidence risk of FLUTD in the DCF- group was
5 ÷ 3354 = 1.49 cases per 1000. The incidence risk of FLUTD in DCF+ cats was 4.01 times greater than the incidence
risk of FLUTD in DCF- cats.

Odds ratio — cohort studies

In a cohort study definition of exposure status (exposure-positive, exposure-negative) comes first.

Subjects are then followed over time to determine their outcome status (disease-positive, disease-
negative). The odds of disease in the exposed and unexposed populations are calculated as follows:
Odds of disease in the exposed population: OE+ = a/b
Odds of disease in the non-exposed population: OE− = c/d
It follows then that the odds ratio for a cohort study is the odds of disease in the exposed population
(OE+ ) divided by the odds of disease in the unexposed population (OE− ):

OE+ ad
OR = = (8)
OE− bc
When the number of cases of disease is low relative to the number of non-cases (i.e. the disease
is rare) a will be small relative to b and c will be small relative to d. This means that the odds ratio
provides a very close estimate of the incidence risk ratio:

a/(a + b) a/b ad
RR = ' = = OR (9)
c/(c + d) c/d bc

39
In the FLUTD cohort study the odds of FLUTD in the DCF+ group was 13 ÷ 2163 = 0.0060. The
odds of FLUTD in the DCF- group was 5 ÷ 3349 = 0.0015. The odds of FLUTD in DCF+ cats was
4.03 times greater than the odds of FLUTD in DCF- cats.

Odds ratio — case-control studies

In a case-control study outcome status (disease-positive, disease-negative) is defined first. The

history provided by each subject then provides information about exposure status. Instead of talking
about the odds of disease in the exposed and unexposed groups (as we did when we were working
with data from a cohort study) we talk about the odds of exposure in the case and control groups:
Odds of exposure in cases: OD+ = a/c
Odds of exposure in controls: OD− = b/d
The odds ratio is defined as the odds of exposure in the cases (OD+ ) divided by the odds of exposure
in the controls (OD− ):

OD+ ad
OR = = (10)
OD− bc

A case-control study was conducted to determine the effect (if any) of the use of CIDR devises on the risk of pregnancy
in dairy cattle. A total of 53 services applied to CIDR-induced oestrus events. Of these 53 services, 23 resulted in
conception. There were 124 services applied to natural oestrus events. Of these 124 services, 71 resulted in conception.
In cases (i.e. cows that conceived) the odds of exposure to a CIDR device was 23 ÷ 71 = 0.32. In controls (cows that
didn’t conceive) the odds of exposure to a CIDR device was 30 ÷ 53 = 0.57. The odds of exposure in cases was 0.32
÷ 0.57 = 0.57 times that of the odds of exposure in controls.

Even though we talk about the odds of exposure in a case-control study, the numeric estimate of the
odds ratio is exactly the same as that calculated for a cohort study. The expression of the result is
the only thing that differs. In a cohort tudy we talk about the odds of disease being x times greater
(or less) in the exposed, compared with the unexposed. In a case-control study we talk about the
odds of exposure being x times greater (or less) in cases, compared with controls.

4.2 Measures of effect in the exposed

Attributable risk (rate)

Attributable risk (or rate) is defined as the increase or decrease in the risk (or rate) of disease in
the exposed group that is attributable to exposure. Attributable risk (unlike incidence risk ratio)
provides a measure of the absolute frequency of disease associated with exposure. Using the
notation defined above, attributable risk (AR) is calculated as:

AR = RE+ − RE− (11)

In a clinical setting attributable risk may also be referred to as attributable risk reduction (ARR) or
attributable risk increase (ARI) depending on whether the event risk is decreased or increased in the
exposure positive group.

40
Figure 17: Explanation of attributable risk as it relates to the FLUTD cohort study. Attributable risk equals the incidence
risk of FLUTD in the exposed (DCF+ cats) minus the incidence risk of FLUTD in the unexposed (DCF- cats), equivalent
to 4.5 cases of FLUTD per 1000 (AR = 0.0045).

A useful way of expressing attributable risk in a clinical setting is in terms of the number needed
to treat (NNT). The NNT is the number of subjects who would have to be given the exposure (e.g.
treatment) to prevent a negative outcome from occurring. NNT equals the inverse of the attributable
risk.
A prospective cohort study was conducted to evaluate the effect of administering oxygen to patients with renal impair-
ment prior to general anaesthesia. The incidence risk of death in oxygen treated patients was 3.5 cases per 100. The
incidence risk of death in patients not receiving oxygen was 6.7 cases per 100. The attributable risk was 3.5 - 6.7 =
-3.2 cases per 100. In other words, oxygen treatment prevented death in 3.2% of patients. The NNT for these data was
-31.3. This means that around 31 patients would need to be treated with oxygen to prevent one death.
NNT gives a good intuitive feel for the treatment benefit and is often useful when communicating the likely effect of
treatment to clients.

Avoid expressing atributable risk (or rate) estimates as a percentage because they are too easily misinterpreted. Always
report them as absolute units, for example the number of cases of disease per 100 head of population.

Attributable fraction

The attributable fraction (also known as the attributable proportion in exposed subjects) is the pro-
portion of disease in the exposed group that is due to exposure. Using the notation defined above,
attributable fraction (AF) is calculated as:

41
 (RE+ − RE− ) (RR − 1)
AF = = (12)
RE+ RR

For case-control studies, attributable fraction can be estimated if the incidence of disease is low:

(OE+ − OE− ) (OR − 1)

AFest = = (13)
OE+ OR

Figure 18: Explanation of attributable fraction as it relates to the FLUTD cohort study. Attributable fraction is the proportion
of incidence risk in the exposed group that is attributable to exposure. In DCF+ cats 75% of FLUTD was attributable to
DCF (AF = 0.75).

In vaccine trials, vaccine efficacy is defined as the proportion of disease prevented by the vaccine in vaccinated indi-
viduals (equivalent to the proportion of disease in unvaccinated individuals due to not being vaccinated), which is the
attributable fraction. A case-control study investigating the effect of oral vaccination on the presence or absence of
rabies in foxes was conducted. Eighteen of 48 unvaccinated foxes developed rabies following challenge. Twelve of 58
vaccinated foxes developed rabies following challenge. The odds of rabies in the unvaccinated group was 2.3 times the
odds of rabies in the vaccinated group (OR = 2.30). Fifty six percent of rabies cases in unvaccinated foxes was due to
not being vaccinated (AFest = 0.56).

42
4.3 Measures of effect in the population

Population attributable risk (rate)

Population attributable risk (or rate) is the increase or decrease in incidence risk (or rate) of disease
in the population that is attributable to exposure. Using the notation defined above, population
attributable risk (PAR) is calculated as:
Incidence risk of disease in the non-exposed population: RE− = c/(c + d)
Incidence risk of disease in the total population: Rtotal = (a + c)/(a + b + c + d)

P AR = Rtotal − RE− (14)

Figure 19: Explanation of population attributable risk as it relates to the FLUTD cohort study. Population attributable risk
equals the risk in the total population minus the risk in the unexposed (DCF- cats). The incidence risk of FLUTD in the cat
population that may be attributed to DCF was 1.8 per 1000. That is, we would expect the risk of FLUTD to decrease by
1.8 cases per 1000 if DCF was not fed (PAR = 0.0018).

Avoid expressing population atributable risk (or rate) estimates as a percentage because they are too easily misinter-
preted. Always report them as absolute units, for example the number of cases of disease per 100 head of population.

Population attributable fraction

Population attributable fraction (also known as the aetiologic fraction) is the proportion of disease in
the study population that is due to the exposure. Using the notation defined above, the population

43
attributable fraction (PAF) is calculated as:
Incidence risk of disease in the non-exposed population: RE− = c/(c + d)
Incidence risk of disease in the total population: Rtotal = (a + c)/(a + b + c + d)

(Rtotal − RE− )
P AF = (15)
Rtotal

Figure 20: Explanation of population attributable fraction as it relates to the FLUTD cohort study. Population attributable
fraction is the proportion of risk in the total study population that is attributable to exposure. Fifty-four percent of FLUTD
cases in the cat population was attributable to DCF (PAF = 0.54).

The population attributable fraction represents the proportion of disease occurrence that would be
eliminated in the population if the exposure is eliminated. Methods are available to estimate PAF
using data from case-control studies.

4.4 Summary

Table 13 outlines which measures of effect are appropriate for each of the three major study designs
(case-control, cohort and cross-sectional studies).
Members of the public often have a poor understanding of relative and absolute risk. A case in point was a recent news
item describing the results of a study of risk factors for leukaemia in children (Draper et al. 2005). Children who lived
within 200 metres of high voltage lines at birth had a 70% higher incidence risk of leukaemia compared with those that
lived 600 metres or more away. While the facts were correctly reported, the interpretation of the scientific evidence was
misguided. If the incidence risk of leukaemia in the general population is around 1 in 20,000 a 70% increase elevates
this to around 2 cases per 20,000 (a very minor increase in absolute terms).

44
 Table 13: Epidemiologic measures of association for independent proportions in 2 × 2 tables.

Parameter Case-control Cohort Cross-sectional

Measures of strength:
Incidence risk ratio No Yes Yes (prevalence RR)
Incidence rate ratio No Yes No
Odds ratio Yes Yes Yes (prevalence OR)

Measures of effect:
Attributable risk No Yes Yes
Attributable fraction No Yes Yes
Attributable fraction (est) Yes Yes Yes

Measures of effect in population (total effect):

Population attributable risk No Yesa Yes
Population attributable fraction No Yesa Yes
Population attributable fraction (est) Yes Yes Yes

a If an estimate of the prevalence of exposure or disease incidence for the population is available from another source.

Figure 21: Newspaper headline warning of the risk of leukaemia associated with living close to high-voltage electricity
lines. Source: The Dominion Post (Wellington, New Zealand) Saturday 4 June 2005.

45
5 Error in epidemiological research

By the end of this unit you should be able to:

• Explain the difference between random error and bias and how each can affect the results of epidemiologic
research.
• Describe the key features of selection and misclassification bias. Provide examples of selection and misclassifi-
cation bias. Explain how you might minimise each type of bias when conducting an observational study.
• Describe the common sources of bias in each of the major epidemiological study designs (ecological, cross-
sectional, cohort, and case-control studies).
• Explain the difference between confounding and interaction, with examples.

5.1 Accuracy and precision

When you derive an estimate of a variable from a sample you want it to be precise and accurate.
A precise estimate has confidence intervals that are small. An accurate estimate has confidence
intervals that are centred on the true population value. There are two types of error that can occur
in epidemiologic research: random error and bias. The difference between random error and bias is
explained in Figure 22.

Figure 22: The bullets fired at the target on the left show little evidence of random error and bias. The bullets fired at the
centre target show a high degree of random error and a low degree of bias. The bullets fired at the target on the right
show a low degree of random error and a high degree of bias.

5.2 Bias

Bias is caused by systematic error. A systematic error is one that is inherent to the measurement
technique being used. It results in a predictable and repeatable error for each observation. Bias
results in observed effect estimates which differ from those which truly exist in the target population
(other than differences due to random error). There are two broad types of bias: selection bias and
misclassification bias. Some authors consider confounding to be a type of bias. In this discussion,
we consider it as a distinct entity as a source of error in epidemiological research.

Selection bias

Selection bias is caused by the procedures used to select units that are included in a study. Selection
bias occurs when these procedures give an observed measure of effect for study participants which
is different to that of non-participants. The different types of selection bias include:

46
 • Surveillance bias: if disease is asymptomatic or mild, it is more likely to be detected in persons
under frequent medical surveillance.

• Referral bias: differential referral patterns are a source of bias in hospital-based case-control
studies.

• Non-response bias: non-response or refusal to participate in a study.

• Length of stay bias: for hospital-based case-control studies, cases and controls should ideally
be selected by a scheme that is equivalent to sampling admission logs (incident cases) rather
than the hospital register of current patients (prevalent cases).

• Survival bias: the introduction of insulin has increased the lifespan of diabetic patients, pro-
ducing an apparent increase in the prevalence of the disease.

Selection bias is usually cited as an issue for observational studies where associations between
exposure and disease are being investigated. However, in descriptive studies (for example, where
the frequency of disease is being described for a given population), the term selection bias can
also be used to describe the situation when the frequency of disease in the group studied is not
representative of that in the target population. This might occur when the sampling frame is not
representative of the target population, and/or when response rates and/or withdrawal rates are
high. Note that, in descriptive studies, the aim is to estimate disease frequency rather than to
investigate associations between exposure and disease. So, in this case, selection bias will result in
disease frequency estimates that differ from that in the target population. Selection bias cannot be
eliminated (controlled for) using analytical techniques. The basic options for avoiding selection bias
are as follows:

• Ensure that study participants are selected at random from the eligible population. Random
selection does not ensure that the study population is representative of the eligible population.
Rather, random sampling allows the probability of differences between the study population
and the eligible population to be assessed. The probability of substantial differences between
the two is greatest when sample sizes are small.

• Ensure that response rates are high amongst the study population.

• Ensure that withdrawal rates are low amongst the study population.

• In observational studies, consider the ‘forces’ which result in individuals being selected.

Misclassification bias

Misclassification (information) bias is due to errors in the information that is recorded for study par-
ticipants. The different types of misclassification bias include:

• Recall bias: cases are generally better at recalling past exposure events, compared with non-
cases.

47
 • Interviewer bias: a potential problem when interviewers are privy to the hypothesis under
investigation.

• Prevarification bias: subjects in a study may have ulterior motives for deliberately overestimat-
ing exposure to a hypothesised causal agent (e.g. compensation claims).

• Improper analysis bias: if one matches cases and controls on a variable that is associated
with the study exposure, an analysis that ignores the matching will yield a disease exposure
odds ratio that is biased towards unity.

• Obsequiousness bias (the ‘Clever Hans’ effect): occurs when subjects systematically alter
their responses in the direction they perceive to be desired by the investigator. Named after
a trained horse who could apparently perform simple arithmetic. Han’s ability stemmed from
nonverbal clues from his trainer that helped him to determine when to stop stamping his hoof
in response to a question.

Misclassification bias can be differential or non-differential. Errors on one axis (e.g. exposure)
that are independent of the other axis (disease) result in non-differential misclassification. Non-
differential misclassification results in the observed measure of association being biased towards
the null. When the measurement error and resulting misclassification occur to a greater extent in
one group than in another they are described as being differential. The effects of differential misclas-
sification are generally harder to predict than those of non-differential misclassification. Differential
misclassification can be very difficult to deal with because, unless you have some idea of how much
is occurring and where it is occurring, you cannot estimate the magnitude of its effect or the direction
of its effect (i.e. whether it shifts the observed measure of association away from, or towards the
null).
Note that the foregoing discussion applies only to studies that ascertain exposure and disease sta-
tus for individuals within a population. In ecological studies (in which the effect of an exposure is
estimated by correlating disease rates across groups of individuals) non-differential misclassification
of exposure can actually lead to an inflated estimate of the influence of exposure on disease risk.
See Brenner et al. (1992) for further discussion of this topic.
Misclassification bias cannot be controlled for using analytical techniques. Options for avoiding
misclassification bias are as follows:

• Ensure that exposure and disease status are assessed independently — i.e. assess one
without knowledge of the other.

• Use a rigorous and biologically valid method for determining the presence of disease and
exposure.

• Use complete and detailed sources of information (i.e. complete exposure histories).

• Use objective measures where available (e.g. liveweights, fleece weights, milk production
records, pregnancy testing results, laboratory measurements).

We usually assume that laboratory measurements apply most frequently to classification of disease
status. However, laboratory measurements can also be useful for determining exposure status. For

48
example, if the exposure of interest was contamination with a chemical, laboratory estimates of
residues may be useful. Trace element, macromineral and parasitological status of livestock may
be assessed using laboratory tests when we are interested in the effects of these exposures rather
than the determinants of these conditions.

5.3 Confounding

Table 14 provides data on mortality rates for six countries in North and Central America for 1986.
A clear trend is evident: mortality rates are relatively low for Central American countries and high
for North American countries such as Canada and the USA. This leads to a question: is there
something special about living in Central America that lowers the overall risk of death? The answer
to this question of course is no, and the reason for the marked difference in mortality is due to the
age distribution of each country’s population. In the Central American countries the population is, on
the whole, younger than the population in the USA and Canada and, for this reason overall mortality
rates are lower.
Table 14: Mortality rates in six North and Central American countries, 1986.

Country Mortality a
Costa Rica 4.0
Venezuela 4.4
Mexico 4.9
Cuba 6.7
Canada 7.3
USA 8.0

a Mortality rate (deaths per 1,000 person years).

In this example, age is said to confound the relationship between country and mortality rate. Con-
founding refers to the distortion of the true underlying relationship between an exposure and an
outcome of interest, because of the influence of a third factor. In this example the true relationship
between country and mortality has been distorted because of the influence of age. Let’s now inves-
tigate this data a little deeper to get a better understanding around how confounding works. Imagine
we’ve been provided with detailed information from the USA and Costa Rica allowing us to calculate,
for each country, mortality rates for the ‘young’ and the ‘old’ (with 50 years being an arbitrary cutpoint
for young and old). The data are shown in Table 15.
There is something strange about Table 15. When stratified by age group mortality rates in the USA
are lower than that of Costa Rica. For young people in the USA mortality rate is 1 per 1000 whereas
in Costa Rica it is 2 per 1000; for old people in the USA mortality rate is 9 per 1000, in Costa Rica
it is 20 per 1000. What is confusing is that when we consider both young and old people together
we see the reverse trend. The USA has a much high mortality rate (8 per 1000 person-years)
compared with Costa Rica (4 per 1000 person years). This ‘distortion’ has come about because the
age distributions of the two populations are vastly different — predominantly young people in Costa

49
 Table 15: Mortality rates for the USA and Costa Rica, stratified by age group.

Age group USA Costa Rica

n deaths n in group Mortality a n deaths n in group Mortality a
Young 1 1000 1 20 10,000 2
Old 90 10,000 9 20 1000 20
Total 91 11,000 8 40 11,000 4

a Mortality rate (deaths per 1,000 person years).

Rica and predominantly older people in the USA. Age is a classic confounder in this example — it
has distorted the true relationship between country and mortality rate.
What we’ve just provided is a conceptual definition of confounding. In epidemiological research
there will be many times when you will need to make a decision whether or not a given variable is
likely to be a confounder. In these situations it is helpful think of the relationship between exposure,
outcome and confounder in the format of Figure 23 and to use the following criteria to assist your
decision making:

1. A confounder must be causally associated with the outcome; and

2. A confounder must be causally or noncausally associated with the exposure; and

3. The confounder and the exposure must be on two separate causal pathways to the outcome.

Figure 23: Schematic diagram of the relationship between an exposure, outcome, and confounder. A unidirectional arrow
indicates the association is causal; a bidirectional arrow indicates a noncausal association.

50
We’re interested in the relationship between smoking and laryngeal cancer. A concern has been raised that drinking
alcohol might act as a confounder in this relationship (i.e. we’re worried that many people who smoke will also be
drinkers). Let’s apply the three criteria:

1. Is drinking a cause of laryngeal cancer? Yes.

2. Is drinking associated with smoking? Yes, the two activities often go together, but taking part in one activity (e.g.
smoking) doesn’t ‘cause’ us to take part in the other (e.g. drinking).
3. Is the association between drinking and laryngeal cancer and the association between smoking and laryngeal
cancer on two separate causal pathways? Yes. Cigarettes cause laryngeal cancer through the effect of carcino-
gens. Alcohol causes laryngeal cancer by acting as a physical irritant on the laryngeal mucosa.

We conclude that drinking is likely to confound the relationship between smoking and laryngeal cancer.

Think of the arrows that connect exposure, outcome and confounder in Figure 23 as water pipes.
The direction of the arrow (which represents the association between one variable and another) can
be thought of in terms of a flow of water and the strength of the association can be thought of in terms
of the water pressure travelling through the pipe. Our interest is to determine the water pressure in
the pipe linking exposure and outcome (i.e. the strength of the association between exposure and
outcome). When there is no confounding, water from the exposure arrives at the outcome directly
(at a given pressure). When confounding is present, water from exposure arrives at the outcome
from two sources: directly from the exposure and via the confounder. In this case, the presence of
the confounder changes the strength of association between the exposure and outcome (changing
the water pressure in the pipe from connecting the exposure with the outcome).
Once we have determined that a variable is likely to be a confounder, we need to consider the likely
direction of its effect. Does the presence of the confounder strengthen or weaken the observed
association? To answer this question, take the following four-step approach.
Step 1. Construct a 2 × 2 table with the two levels of exposure as rows and the two levels of disease
status as columns. Figure 24 shows a 2 × 2 table constructed to evaluate the association between
smoking and cancer of the larynx.

Figure 24: 2 × 2 table with smoking (rows) as the exposure and laryngeal cancer (columns) as the outcome.

Step 2. Think about the effect of the confounder on each level of exposure. In this example, we ask
the question: who are bigger drinkers — smokers or non-smokers? Answer: drinking is likely to be
positively associated with smoking. Indicate this effect on your 2 × 2 table, as shown in Figure 25.
Step 3. Think about the effect of the confounder on each level of the outcome. Is drinking likely to be
positively or negatively associated with laryngeal cancer? Answer: drinking is likely to be positively

51
Figure 25: 2 × 2 table for the association between smoking and laryngeal cancer showing the likely effect of drinking on
smoking.

associated with laryngeal cancer. Again, indicate this effect on your 2 × 2 table, as shown in Figure
26.

Figure 26: 2 × 2 table for the association between smoking and laryngeal cancer showing the likely effect of drinking on
laryngeal cancer.

Step 4. Interpret the results. Drinking is positively associated with smoking, so we anticipate an
overrepresentation of subjects in the exposure-positive (a and b) cells of the 2 × 2 table. Drinking
is positively associated with laryngeal cancer, so we anticipate an overrepresentation of subjects in
the outcome-positive (a and c) cells of the 2 × 2 table. Based on these judgements, the size of the
a cell will be exaggerated. This means that the proportion of smokers with laryngeal cancer will be
increased, resulting in a strengthening of the primary exposure-outcome association.
We conclude that drinking confounds the association between smoking and laryngeal cancer, mak-
ing the association between smoking and laryngeal cancer seem stronger than it actually is.

5.4 Interaction

Confounding is said to be present when the summary measure of association is misrepresented

because of the presence of a third (confounding) variable. Interaction — also known as effect modi-
fication — is said to be present when the strength of association between an exposure and a given
outcome varies with the level of a second exposure variable.

52
Examples of interaction:

• Drinking and tranquilisers. If we drink and take tranquilisers at the same time, the effect on cognitative ability will
be greater than the sum of the individual effects of alcohol and tranquilisers taken alone.
• The relationship between salt consumption and the risk of stroke is quite different for men and women. Women
with high salt intakes have a moderately increased risk of stroke than do other women. Men with high salt intakes
have a substantially greated risk of stroke than do other men.

There are two types of interaction: positive (synergism) and negative (antagonism). With positive
interaction the observed joint effect of two factors is greater than that expected by summing their
independent effects. With negative interaction the observed joint effect of two factors is less than that
expected by summing their independent effects. Figure 27 presents these concepts diagramatically.

53
 (a) No interaction: the observed joint effect of factors A and B equals the
sum of their independent effects.

(b) Positive interaction: the observed joint effect of factors A and B is greater
than the sum of their independent effects.

(c) Negative interaction: the observed joint effect of factors A and B is less
than the sum of their independent effects.

Figure 27: Diagram illustrating the concepts of no, positive, and negative interaction.

54
We’ll now work through an example in more detail. Milk fever (hypocalcaemia) is a risk factor for
conception failure in dairy cattle. We want to know if this association is stronger in older cattle, com-
pared with younger cattle. That is, does the association between milk fever and failure to conceive
interact with age? The data shown in Table 16 were collected.
Table 16: Milk fever as a risk factor for failure to conceive in dairy cattle, stratified by age. Additive interaction absent.

Age Milk fever Incidence AR (%)

Young Absent 10%
Young Present 20% 10%
Old Absent 20%
Old Present 30% 10%

In Table 16 the presence of milk fever increases the incidence risk of failure to conceive by 10% in
both young and old cows. Because the attributable risks associated with milk fever (the exposure)
are not modified by age, we conclude that additive interaction is absent: the attributable risks are
the same for each level of age. Now consider an alternative scenario, shown in Table 17.
Table 17: Milk fever as a risk factor for failure to conceive in dairy cattle, stratified by age. Additive interaction present.

Age Milk fever Incidence AR (%)

Young Absent 5%
Young Present 10% 5%
Old Absent 10%
Old Present 30% 20%

In Table 17 the attributable risks associated with milk fever are modified by age, so we conclude that
additive interaction is present. The interaction plots shown in Figure 28 illustrate this effect.
The data shown in Table 17 is an example of additive interaction. We now consider multiplicative
interaction. Hypothetical data showing the association between milk fever and failure to conceive for
two levels of age are shown in Table 18. This time, instead of expressing the strength of association
in terms of attributable risk, we express it in terms of the incidence risk ratio.
Table 18: Milk fever as a risk factor for failure to conceive in dairy cattle, stratified by age. Multiplicative interaction absent.

Age Milk fever Incidence Risk ratio

Young Absent 10%
Young Present 20% 2.0
Old Absent 20%
Old Present 40% 2.0

Because the incidence risk ratios for milk fever and failure to conceive are not modified by age
we conclude that multiplicative interaction is absent. Table 19 shows how the data might look if

55
Figure 28: Additive interaction. The plots above are based on the data presented in Tables 16 and 17. The plot on the
left shows the situation where additive interaction is absent (Table 16): the attributable risk for each level of age are the
same. The plot on the right shows the situation where additive interaction is present (Table 17): the attributable risk for
older cows (20%) is greater than that for younger cows (10%).

multiplicative interaction was actually present. For young cattle with milk fever the risk of failure to
conceive is increased by a factor of 2. For old cattle with milk fever the risk of failure to conceive
is increased by a factor of 3. Interaction plots for the data shown in Tables 18 and 19 are shown in
Figure 29. Note that in Figure 29 the vertical axis of each plot is shown on the logarithmic scale.

Table 19: Milk fever as a risk factor for failure to conceive in dairy cattle, stratified by age. Multiplicative interaction present.

Age Milk fever Incidence Risk ratio

Young Absent 10%
Young Present 20% 2.0
Old Absent 20%
Old Present 60% 3.0

A second strategy for identifying the presence of interaction is to compare the joint observed and
joint expected effects. In Table 20 we show the same data presented in Table 16 including the
attributable risk for each strata less the attributable risk for the reference group (young cows that are
milk fever negative). With this done, we then calculate the joint observed and expected attributable
risks.
Joint observed attributable risk: Obs AR[Age+M F +] = 20%.
Joint expected attributable risk: Obs AR[Age−M F +] + Obs AR[Age+M F −] = 10% + 10% = 20%.
Because the joint observed attributable risk is the same as that expected by adding the individual
attributable risks we conclude that additive interaction is absent. In Table 21 we repeat the analysis
for the situation where additive interaction is present, using the data shown in Table 17.
Joint observed attributable risk: Obs AR[Age+M F +] = 25%.
Joint expected attributable risk: Obs AR[Age−M F +] + Obs AR[Age+M F −] = 5% + 5% = 10%.

56
Figure 29: Multiplicative interaction. The plots above are based on the data presented in Tables 18 and 19. The plot on
the left shows the situation where multiplicative interaction is absent: the incidence risk ratio for each level of age are the
same. The plot on the right shows the situation where multiplicative interaction is present: the incidence risk ratio for older
cows (3.0) is greater than that for younger cows (2.0). Note the logarithmic scale of the vertical axes.

Table 20: Milk fever as a risk factor for failure to conceive in dairy cattle and the interacting effect of age. Comparison of
joint observed and expected attributable risks — additive interaction absent.

Age Milk fever Incidence AR (%) Obs AR (%)

Young Absent 10% -
Young Present 20% 10% 10%
Old Absent 20% 10%
Old Present 30% 10% 20%

Because the joint observed attributable risk is different to that expected by adding the individual
attributable risks we conclude that additive interaction is present.
Why is it important to distinguish between additive and multiplicative interaction? Many of the com-
mon regression techniques used to control for the effect of confounders (e.g. logistic regression)
assume that interaction is multiplicative whereas most interactions in biology are additive. To evalu-
ate additive interaction appropriately in a multivariable model, redefine the two exposures in question
by considering them jointly as a single composite exposure variable and entering combinations of
exposures into the model as a factored set of terms. For two dichotomous exposures (A and B)
the composite variable would have four levels: A-B-, A+B-, A-B+, and A+B+. Using A-B- as the
reference category, the model will provide estimates of the relative effect for each of the other three
categories. This approach allows you to assess departure from additivity, without imposing the mul-
tiplicative relation implied by the regression technique being used. An excellent discussion of the
issues around additive and multiplicative interaction is provided by Rothman (2002, pp 168 – 180)
and Greenland (2009). A nice paper covering some of the technical aspects of additive interaction
in logistic regression models is provided by Knol et al. (2007).

57
Table 21: Milk fever as a risk factor for failure to conceive in dairy cattle and the interacting effect of age. Comparison of
joint observed and expected attributable risks — additive interaction present.

Age Milk fever Incidence AR (%) Obs AR (%)

Young Absent 5% -
Young Present 10% 5% 5%
Old Absent 10% 5%
Old Present 30% 20% 25%

5.5 Dealing with confounding and interaction

Confounding and interaction are different phenomena and different strategies are used to identify
their presence in an exposure-outcome data set. In the first instance, when you suspect that a
variable is confounding or interacting with an exposure-outcome relationship, stratify the data by
the variable to calculate strata-level measures of association. If the measures of association are
significantly different among strata, then the conclusion is that interaction is present and you should
report the measures of association for each strata.
If the strata-level measures of association are not significantly different, interaction is said to be
absent and we move on to assess for the presence of confounding. The first step when determining
whether or not a variable is a confounder is to apply the three criteria listed above: (1) is the variable
causally associated with the outcome? (2) is the variable noncausally associated with the exposure?
and (3) is the variable and the exposure on two separate causal pathways to the outcome? The final
step is to apply some quantitative criteria to determine if the confounder is having a detectable
influence on the exposure-outcome relationship under investigation. What we do here is calculate a
measure of association that adjusts for the effect of the confounder (e.g. the Mantel-Haenszel risk
ratio) and compare the adjusted measure with the crude measure of association. A rule of thumb is
that if the adjusted measure of association differs from the crude measure of association by 10% to
15% then we conclude that confounding is present at a level that is sufficient to warrant correcting
for it in our analyses.
If there are i strata and the total number of subjects in each strata is Ti, the crude odds ratio is
calculated using Equation 16.

X X
ai di
i i
ORcrude = X X (16)
bi ci
i i

The Mantel-Haenszel adjusted odds ratio is calculated using Equation 17.

X ai di
Ti
i
ORM-H = X (17)
bi ci
Ti
i

58
The formula for the Mantel-Haenszel adjusted odds ratio is provided here simply to give you an
idea of what the adjustment process involves. Formulae for adjusting other measures of association
are provided in many standard epidemiological texts. Elwood (2007) provides a very readable and
clear description of the approach. A summary of the methodological approach for distinguishing
confounding and interaction in a data set is shown in Figure 30.

Figure 30: Strategy for distinguishing confounding and interaction.

A confounder is an extraneous factor that wholly or partially accounts for the observed effect of a risk factor on dis-
ease. Interaction, as distinct from confounding, occurs when the effect of an exposure on an outcome depends on the
presence or absence of a third factor.
Interaction results in measures of association that differ across strata. Because the strata-level measures of association
differ, summary measures of association are not appropriate.
Confounding results in measures of association that are the same across strata and summary (adjusted) measures of
association differ from the crude (unadjusted) measures of association.

Methods for dealing with confounding

Restriction
Think about the association between the number of children a woman has had and the risk of being
diagnosed with breast cancer and the confounding effect of age. To deal with age as a confounder in
the relationship between parity and breast cancer we could only include those subjects that were of
a certain age in the study population. We could do this with either a cohort or a case-control design.
Restriction is clearly an effective method, as it leaves no possibility of confounding, but obviously
the disadvantage is that the study then becomes specific to a particular age group, and we cannot
generalise the study beyond that target population.

59
Randomisation
The principle of randomisation is that from a pool of study participants, subjects are randomly as-
signed to exposure and non-exposure groups. The definition of random is such that each subject in
the study has the same chance of being allocated to a particular group, and that the chance of one
individual being allocated to one group is not influenced by the allocation of any other member to
the group. The advantage of randomisation is that, given large sample sizes, it is likely to produce
groups which are similar even in respect to variables which have not been anticipated, designed, or
measured.
Randomisation is an option for dealing with confounding in prospective intervention studies assess-
ing the effects of an ethical, practical and acceptable intervention which is thought to be beneficial
and not likely to be harmful. Randomisation cannot be applied to retrospective studies.
Stratification
The best way to adjust for a single confounder is to examine exposure-outcome relationships within
levels of the confounder. Within each of the confounder levels, there will be no confounding because
exposed and non-exposed subjects will all have the same level of the confounder. If the size of the
exposure-outcome association is the same at all levels (or strata) of the confounder, then statistical
methods can be used to combine the stratum-specific estimates of effect to give an estimate of effect
that is adjusted for the confounder.
Matching
Matching each exposed subject to an unexposed subject with the same level of a confounder will
reduce selection bias. For example, in the hypothetical smoking/heart disease study, smokers and
non-smokers could be matched according to sex. When a male smoker is recruited into the study,
he is matched with a male non-smoker. When a female smoker is recruited she is matched to a
female non-smoker. This will obviously lead to identical percentages of men and women among
smokers and non-smokers.
Matching of exposed and non-exposed subjects is only possible in studies where subjects are re-
cruited on the basis of their exposure status. It is not possible in case-control studies, where subjects
are recruited according to their outcome status (presence or absence of outcome). Thus, in a case-
control study of smoking and coronary heart disease, people with heart disease could be matched
by sex to people without heart disease. This would result in cases and controls having the same per-
centage of male and females but would not lead to an even distribution of sex among smokers and
non-smokers. The latter condition is the important one for control of confounding by sex. Matching is
an excellent design strategy for control of confounders in cohort studies. However, it is inappropriate
for this purpose in case-control studies.
The purpose of matching in case-control studies is to improve the statistical power of the study. If
matching is done to improve power in a case-control study, then the data analysis should take this
matching into account. Otherwise, bias can be introduced into the study.
Multivariate methods
Whereas stratification is an excellent method for controlling a single confounder, multivariate meth-
ods (statistical modeling) is required if there are multiple confounders. One disadvantage of mod-
eling as a means to control confounding is that the investigator is distanced from the mechanics of
the data analysis: stratification permits a much better ‘feel’ for the data and should always precede
modeling.

60
A worked example

Siscovick et al. (1984) conducted a case-control study to evaluate the relationship between primary
cardiac arrest and habitual vigorous exercise. They reported the data shown in Table 22. We would
like to know: (1) if there is an interaction between smoking and vigorous habitual exercise on the risk
of primary cardiac arrest, and (2) if smoking confounds the association between vigorous habitual
exercise and primary cardiac arrest. First, we check for evidence of interaction.

Table 22: Results of a case-control study evaluating the relationship between habitual exercise and primary cardiac arrest
(Siscovick et al. 1984).

Non-smokers Diseased Non-diseased Total

Exercise - 36 24 60
Exercise + 32 70 102
Total 68 94 162

Smokers Diseased Non-diseased Total

Exercise - 40 17 57
Exercise + 25 22 47
Total 65 39 104

Figure 31: Interaction plot showing the nature of the interaction between smoking and habitual vigorous exercise on the
risk of primary cardiac arrest.

The odds of primary cardiac arrest for non-smokers that did not undertake habitual vigorous exercise
was 3.28 (95% CI 1.69 – 6.38) times that of those who did exercise habitually. The odds of primary
cardiac arrest for smokers that did not undertake habitual vigorous exercise was 2.07 (95% CI 0.92
– 4.64) times that of those who did exercise habitually. There appears to be a synergistic interaction
between smoking, exercise, and risk of primary cardiac arrest (Figure 31).

61
Although we have evidence to suggest the presence of interaction, we need to test the hypothesis
that the strata-level odds ratios are the same using the chi-squared test of homogeneity. The test of
homogeneity test statistic is compared with a chi-squared distribution with n - 1 degrees of freedom
(where n is the number of strata). A test of homogeneity of the stratified odds ratios produces a χ2
test statistic of 1.03. Since there are two strata, the comparison has 1 degrees of freedom and the
associated P-value is 0.31. We accept the null hypothesis and conclude that the stratum-specific
odds ratios are the same (that is, there is no significant interaction). We now apply the three criteria
outlined at the start of this section to determine if smoking confounds the relationship between
habitual vigorous exercise and primary cardiac arrest.
Is smoking causally associated with primary cardiac arrest?

Diseased Non-diseased Total

Smoking + 65 39 104
Smoking - 68 194 162
Total 133 133 266

The odds of primary cardiac arrest for smokers was 4.75 (95% CI 2.93 – 7.71) times that of non
smokers. We have evidence that smoking is associated with cardiac arrest. A review of the relevant
literature would also support the notion that this association is causal.
Is smoking noncausally associated with habitual exercise?

Exercise+ Exercise - Total

Smoking + 47 57 104
Smoking - 102 60 162
Total 149 117 266

The odds of being a habitual exercise for smokers was 0.49 (95% CI 0.29 – 0.80) times that of non
smokers. It is reasonable to conclude that being a smoker is noncausally associated with (lack of)
habitual exercise.
Is the link between habitual exercise and cardiac arrest, and smoking and cardiac arrest on two
separate causal pathways? Lack of habitual exercise and smoking increase the risk of cardiac
arrest by two independent physiological mechanisms. It is reasonable to assume that they are on
two separate causal pathways.
Does the strength of the association between habitual exercise and primary cardiac arrest change
when you account for the presence of smoking? The crude odds of primary cardiac arrest in those
who undertook habitual vigorous exercise was 2.99 (95% CI 1.81 – 4.95) times greater than those
who did not. The Mantel-Haenszel adjusted odds of primary cardiac arrest for those that did not
undertake habitual vigorous exercise was 2.72 (95% CI 1.46 – 5.06). The ratio of the crude odds
ratio to the adjusted odds ratio is 2.99 ÷ 2.72 = 1.10. We conclude that smoking confounds the
association between habitual vigorous exercise and risk of primary cardiac arrest (using a relative
difference of more than 10% to 15% between the crude and adjusted odds ratio as an objective
indicator of the presence of confounding).

62
6 Causation

By the end of this unit you should be able to:

• Explain the difference between association and causation in epidemiological research.

• Define component, sufficient, and necessary causes (with examples).
• List and briefly explain each of Hill’s criteria for causation.

6.1 Association versus causation

A fundamental objective of epidemiologic research is to identify the causes of disease through the
study of the distribution of cases within groups of individuals with identified characteristics, such as
different levels of exposure to some agent (e.g. exposure to a drug or chemical). Knowledge of what
causes disease allows us to develop prevention strategies by targeting those risk factors influential
in determining the likelihood that disease will occur.
Using this approach we need to be aware of the difference between association and causation.
Association is a quantitative measure of the strength of the relationship between an exposure and
outcome. A cause, on the other hand, is the presence of a combination of exposures that alone or
in combination and in the correct sequence and timing during an animal’s life inevitably result in an
outcome (such as clinical disease).
A study conducted in the 1980s found that dairy herds milked by staff who wore shorts and aprons during milking were
more likely to be positive for leptospirosis. These findings lead to the question: do shorts and plastic aprons cause
leptospirosis, or are they associated with the presence of leptospirosis? Obviously, shorts and aprons are associated
with the presence of leptospirosis, and in this example their use was a marker (i.e. a proxy variable) for other causative
factors, such as (for example) herd size.
Reference: Mackintosh C, Schollum L, Harris R, Blackmore D, Willis A, Cook N, Stoke J (1980). Epidemiology of
leptospirosis in dairy farm workers in the Manawatu. Part I: A cross-sectional serological survey and associated occu-
pational factors. New Zealand Veterinary Journal 28: 245 - 250.

When a previously unrecognised disease is identified epidemiological research usually starts with
case reports and case series that describe the condition and provide evidence that the disease can
occur repeatedly. Descriptive studies follow, where the distribution of disease is documented accord-
ing to individual, place, and time. Descriptive studies are useful because they provide a rich source
of hypotheses about factors that are associated with, or cause the disease. Analytical studies (i.e.
cross-sectional, cohort, case-control studies) provide a means for testing the hypotheses generated
from descriptive studies.
Because epidemiology is predominantly an observational (i.e. non-experimental) science that draws
its data from the uncontrolled conditions, we need to be aware that bias, confounding, and chance
may provide alternative explanations for the associations that we might identify. If these issues are
thought to be present then further analytical studies need to be undertaken to account for them.
Once we are confident of the validity of the associations that have been observed (i.e. we’re con-
fident bias, confounding, and chance are not present) attention turns towards establishing if the
relationships between the identified risk factors and disease are causal. Figure 32 provides an
outline of the process.
Whereas the identification of association is predominantly a quantitative process, identifying causal
relationships is largely subjective and based on judgement. Over the years, several authors (e.g.

63
Koch, Evans, and Hill) have defined criteria that might be used to help identify is an observed
relationship is causal.

Figure 32: Flow diagram outlining the typical sequence of events in the epidemiological investigation of disease.

6.2 Component, sufficient, and necessary causes

Causes have the following characteristics: (1) they must precede the effect, (2) they can be either
host or environmental factors (e.g. characteristics, conditions, actions of individuals, events, natural,
social or economic phenomena), and (3) they can be either positive (the presence of an exposure)
or negative (the absence of exposure, such as vaccination).
The key thing is that one or more causes are determinants of many of the diseases that we deal
with. On one hand we might have a condition such as anthrax, which has a single cause (exposure
to Bacillus anthracis). For other diseases, such as lameness in dairy cattle, there might be multiple
causes (e.g. poor hoof condition, injury, age). This said, it is easiest to conceptualise causation by
regarding causal factors as the pieces of a pie. Disease occurs when we have assembled enough
causal factors to to produce a full pie. For some diseases (especially infectious conditions) it may be
that exposure to the infectious agent will cause disease: in this situation there is only one piece to

64
the pie. For other diseases there may be many reasons why some exposed individuals don’t develop
the disease yet others do: in this situation, the pie is made up of many pieces. The following terms
are used when talking about causation:

• Component causes are conditions that are causally related to the presence of disease (the
pieces of the pie). Factors such as high cholesterol, smoking, lack of exercise, genetics, and
the presence of concurrent diseases are all component causes of coronary heart disease in
humans.

• Sufficient causes are the set of conditions without any one of which disease would not have
occurred (the whole pie). Sufficient causes are not usually a single factor but several. Accu-
mulation of a set of sufficient causes is synonymous with occurrence (although not necessarily
diagnosis) of disease.

• A necessary cause is one that must be present for the disease to occur (the most important
piece of the pie). If chicken salad has been identified as sufficient causes of salmonellosis in
a foodborne disease outbreak, Salmonella spp. would be a necessary cause of diarrhoea.

Figure 33: This diagram represents a disease that has two sufficient causal complexes: the first having three component
causes and the second having two. A, B, C, and D are component causes. A is a necessary component cause.

Whereas infection with Mycobacterium tuberculosis is a necessary cause for tuberculosis disease, it is not a sufficient
cause since many animals may harbor small foci of tuberculosis without developing tuberculosis.
Tobacco smoking is a sufficient cause of lung cancer, but so is exposure to radon or certain occupational chemicals in
non smokers.
Coronary heart disease in humans has no necessary cause, but rather a range of component causes which become
sufficient when some or all occur together in individuals at levels that cumulate and interact to result in disease.

Koch (1884) was the first to provide a framework for identifying causes of infectious disease. Koch
specified that the following criteria (known as Koch’s postulates) had to be met before an agent could
be considered as the cause of a disease:

1. The agent has to be present in every case.

2. The agent has to be isolated from the affected individual and grown in pure culture.

3. The agent has to cause disease when inoculated into a susceptible animal and the agent must
then be able to be recovered from that animal and identified.

In the late nineteenth century Koch’s postulates brought a degree of order and discipline to the
study of infectious diseases, although the key assumption of ‘one-agent-one-disease’ was highly

65
restrictive since it failed to take account of diseases with multiple aetiologic factors, multiple effects
of single causes, carrier states, and non-agent factors (such as age and sex). Based on John Stuart
Mill’s rules of inductive reasoning from 1856, Evan developed a unified concept of causation which is
now the generally accepted means for identifying cause-effect relationships in modern epidemiology.
Evan’s unified concept of causation includes the following criteria:

1. The proportion of individuals with disease should be higher in those exposed to the putative
cause than in those not exposed.

2. Exposure to the putative cause should be more common in cases than in those without the
disease.

3. The number of new cases should be higher in those exposed to the putative cause than in
those not exposed, as shown in prospective studies.

4. Temporally, the disease should follow exposure to the putative cause.

5. There should be a measurable biologic spectrum of host responses.

6. The disease should be reproducible experimentally.

7. Preventing or modifying the host response should decrease or eliminate the expression of
disease.

8. Elimination of the putative cause should result in lower incidence of disease.

6.3 Hill’s criteria

Bradford Hill (1965) elaborated on Evan’s criteria as part of work that identified smoking as a cause
of lung cancer. Hill’s criteria are as follows:

1. Strength of association

2. Consistency.

3. Temporality.

4. Dose response relationship.

5. Plausibility and coherence.

6. Experimental evidence.

7. Specificity.

8. Analogy.

Hill’s intention was to provide a set of guidelines that could be used to determine if associations are
causal, providing the following cautionary statement: ‘none of my viewpoints can bring indisputable
evidence for or against the cause and effect hypothesis and none can be regarded as sine qua non
1
.’
1
sine qua non: an essential condition or element

66
Strength of association

The first criterion is that of strength of association, which is conventionally measured by the risk (or
odds) of disease in exposed individuals compared with the risk (or odds) in the unexposed. The
rationale here is that strong associations are unlikely to be a result of uncontrolled bias or confound-
ing. Strong associations are usually considered to be risk ratios in excess of 4 or 5. The term ‘small’
is used in epidemiology to describe risk ratios from observation studies that are less than 2.0, since
it is possible that such associations may be due to bias and/or confounding. Obviously, a relative
risk of 1.4 is not actually small in magnitude since it indicates a 40% higher rate in the exposed,
and would have a significant demographic effect when applied to populations if the exposure was
common. Intervention effects of this magnitude would be regarded as highly significant from a pub-
lic health or clinical perspective if obtained from randomised trials where confounders are dealt with
in the randomisation process. As methods and analysis in observational epidemiology continue to
improve, smaller relative risks may be accepted as evidence of causation.
Establishment of an overall risk ratio from a number of studies can be achieved by an analytical
technique known as meta-analysis, which may vary from the use of mean or median values to fixed-
and random-effects models. Combination of data from several studies may produce a statistically
significant risk ratio, whereas individual studies may lack sufficient numbers to achieve statistical
significance.

Consistency

Consistent findings from several studies that have investigated the strength of association between
a risk factor and disease support an argument that the risk factor is causative. Consistency also
applies to the existence and pattern of trend in dose response.
Smoking has been associated with lung cancer in at least 29 retrospective and 7 prospective studies. The consistency
of this association provides powerful evidence that smoking causes lung cancer.

Temporality

Causes must preceed the effect. For example, if severe angina due to coronary heart disease led
to reduced physical activity and a sedentary lifestyle in a previously active person, such inactivity
(although associated with coronary heart disease in a cross-sectional context) could not be held ac-
countable for it. Longitudinal studies are particularly useful for determining the temporal relationship
between possible causative factors and outcomes.
Obesity can be identified as a strong risk factor for the incidence of adult onset diabetes in longitudinal studies. However,
since the management of adult onset usually results in weight loss, cross-sectional studies of obesity and diabetes may
reveal no association, or even a negative one.

Dose response

A dose response effect implies that the likelihood or severity of the outcome is greater with a higher
close of the exposure. This may manifest as a comparison between outcomes at multiple levels
of exposure. Trends may be linear or curvilinear. While differences between individual exposure

67
groups may not be statistically significant, the trend across three or more groups may be significant.
When several studies are considered it is usual to report what proportion demonstrates a significant
trend, and whether significant trends are in the same direction.

Plausibility and coherence

Here we ask if a causal interpretation fits with known facts of natural history and biology of disease,
that is does a causal relationship make ‘biological sense’? Biological plausibility provides a strong
argument for causation, if it is present. However, its absence need not exclude causality, particularly
if little is known about the disease under investigation.

Experimental evidence

As outlined above, it is not generally possible to perform experiments on humans in which possible
disease-producing agents are administered or risk behaviors encouraged. However, evidence from
controlled randomised trials of interventions provide a good argument for causation.
The beneficial effects of a high fibre diet in the prevention of colon cancer in high-risk populations suggests that low fibre
intake is a causative factor in disease occurrence.
Reversal of signs of scurvy with vitamin C suggests that vitamin C deficiency is the cause of scurvy.

Specificity

This criteria states that a single exposure generally causes a single disease. This is a hold-over
from the concepts of causation that were developed for infectious diseases, though there are many
exceptions (e.g. smoking is associated with lung cancer as well as many other diseases). When
present, specificity provides evidence of causality, but its absence does not preclude causation.

Analogy

This criteria asks if a similar relationship been observed with another exposure and/or disease (e.g.
BSE and scrapie/transmissible mink encephalopathy). This is one of the weakest criteria for causa-
tion but it is useful in speculation of how putative causative factors may operate in different contexts.

6.4 Causal web models

Causal factors act in a hierarchical fashion and for this reason it is useful to develop path models to
describe and explain the relationships between sufficient and necessary causes of disease. Figure
34 is a path web model showing factors associated with pneumonia in lambs.
Path models are useful because they provide a representation of ‘the big picture’ — a framework for
thinking about the relationships between causes (particularly temporal relationships) and how they
interelate with each other. This provides a useful starting point for designing research strategies to
investigate diseases that are of interest.

68
Figure 34: Path model of factors associated with pneumonia in New Zealand lambs. Reproduced from Goodwin-Ray et
al. (2008).

69
7 Sampling

By the end of this unit you should be able to:

• Explain the key features of simple random sampling, systematic random sampling, stratified random sampling,
and cluster sampling. Describe the advantages of disadvantages of each approach.
• Describe ways to reduce error when making inferences from sampled data.
• Given the appropriate formulae, calculate the required sample size when you want to estimate a population total,
mean, or proportion.
• Given the appropriate formula, calculate the sample size required to detect the presence of disease in a popula-
tion. Adjust this estimate to account for a test that is imperfect.

Epidemiologists frequently examine populations to:

• Detect the presence of a disease;

• Demonstrate that a disease is not present within a population; and

• Establish the level of occurrence of a disease within a population.

To produce accurate estimates of disease we must be able to measure populations effectively. The
exact level of disease within a population will be obtained if every individual within the population is
examined (and if there was no measurement error). This technique is a census. However, in many
situations a census is impossible and/or excessively expensive. Usually an accurate estimate can
be obtained by examining some of the animals (a sample) from the population.

7.1 Probability sampling

A probability sample is one in which every element in the population has a known non-zero chance
of being included in the sample.

Simple random sampling

Simple random sampling occurs when each subject in the population has an equal chance of being
selected.

Figure 35: Simple random sampling. If a sample of five cows was required, five random numbers between 1 and 10
would be generated and cows selected on the basis of the generated random numbers.

70
Systematic random sampling

With systematic random sampling, the selection of sampling units occurs at a predefined equal
interval (known as the sampling interval). This process is used when the total number of sampling
units is unknown at the time of sampling (e.g. in a study where patients that enter an emergency
department of a hospital on a given day are to be sampled — at the start of the study day we do not
know the total number of patients seen by the end of the day).
Suppose we are studying inpatient medical records on an ongoing basis for a detailed audit. The total number of records
in the population is not likely to be known in advance of the sampling since the records are to be sampled on an ongoing
basis (and so it would not be possible to use simple random sampling). However, it would be possible to guess the
approximate number of records that would be available per time period and to select a sample of one in every k records
as they become available.
Say we require a total of 300 records over a 12-month period to complete the study. If there are, on average, ten new
discharge records available per day then total number of records available per year is estimated to be 10 × 365 = 3650.
To obtain the required number of records per year in the sample, the sampling interval k will be 3650 ÷ 300 = 12. Thus,
we would take a sample of 1 from every 12 records.
One way to implement this procedure is to identify each record as it is created with a consecutive number. At the
beginning of the study a random number between 1 and 12 is chosen as the starting point. Then, that record and every
twelfth record beyond it is sampled. If the random number chosen is 4, then the records in the sample would be 4, 16,
28, 40, 52, and so on.

Stratified random sampling

Stratified sampling occurs when the sampling frame is divided into groups (strata) and a random
sample taken from within each stratum. Stratified sampling is frequently undertaken to ensure that
there is adequate representation of all groups in the population in the final sample. The simplest
form is proportional stratified random sampling, where the number sampled within each stratum is
proportional to the total number within the stratum.
Suppose that you wish to determine the prevalence of disease in the pig population of a region. Previous surveys have
indicated that 70% of the region’s pigs are located in very large, intensive specialised pig farms, 20% of pigs are found
within smaller farming units (frequently as a secondary enterprise on large dairy farms), and 10% of pigs are kept singly
within small plots around towns (by people whose major occupation is not farming). With proportional stratification, a
sample would be selected at random from within each stratum such that the aggregated sample would consist of 70%
pigs obtained from the large intensive farms, 20% pigs obtained from the smaller pig farms, and 10% pigs obtained from
small plots near towns.
If the population can be divided into logical strata whereby the variation within each stratum is small
compared with the variation between strata, stratified random sampling will provide a more precise
estimate of the population parameter.
We wish to determine average total lactation milk volume (total litres) produced by dairy cows in a region. The region
contains two breeds of cattle. One breed (Friesian) is characterised by production of large volumes of milk with low
concentrations of milk solids. The other breed (Jersey) is characterised by production of small volumes of milk with high
concentrations of milk solids. By dividing the population into breed strata and sampling within each stratum, the average
lactation milk volume production of each breed can be estimated with accuracy. The mean milk production for cows
within the region can be estimated by working out the weighted mean based upon each stratum mean and stratum size.

Cluster sampling

Cluster sampling occurs when the sampling frame is divided into logical aggregations (clusters)
and a random selection of clusters is performed. The individual sampling units (known as primary

71
Figure 36: Stratified random sampling. A group of animals are stratified by breed and a random sample within each breed
taken.

sampling units) within the selected clusters are then examined. Clustering may occur in space or
time. For example, a litter of piglets is a cluster formed within a sow, a herd of dairy cows is a cluster
within a farm, and a fleet of fishing boats is a cluster formed within space (that is, a port or harbour).
The standard errors of population estimates derived from cluster sampling are often high compared
with those obtained from simple random or stratified random sampling procedures. The reason for
this is that units within the same cluster tend to be more homogenous than those units from different
clusters.
There are two types of cluster sampling:

• One stage cluster sampling occurs when clusters are selected by simple random sampling
and then, once selected, all of the listing units within the cluster are examined.

• Two stage cluster sampling occurs when clusters are selected by simple random sampling
and then, once selected, a random sample of listing units within each cluster are selected for
examination. Estimation of population characteristics is straightforward in this situation when
each cluster has the same number of listing units. Estimation of population characteristics is
not straightforward when each cluster contains different numbers of listing units (in this case,
you will need to consult a statistician).

The number of clusters to sample and the number of listing units within each cluster to sample will
depend upon the relative variation of the factor of interest between clusters, compared with within
clusters, and the relative cost of sampling clusters compared with the cost of sampling individual
listing units.

72
 • When the between-cluster variation is large relative to the within-cluster variation, you will have
to sample many more clusters to get a precise population estimate.

• When the between-cluster variation is small relative to the within-cluster variation, you will have
to sample many more individual listing units within each cluster to get a precise population
estimate.

7.2 Non-probability sampling

Non-probability sampling occurs when the probability of selection of an individual within a population
is not known and some groups within the population are more or less likely than other groups to be
selected. Non-probability sampling methods include:

• Convenience sampling: where the most accessible or amenable sampling units are selected;

• Purposive sampling: where the most desired sampling units are selected; and

• Haphazard sampling: where sampling units are selected using no particular scheme or
method. Inherent in this type of sampling is the problem that subconscious forces may in-
fluence the person selecting the units in an attempt to ‘balance’ the sample. For example, a
young animal may be preferred for the next selection immediately after an older animal has
been selected.

Non-probability sampling will produce biased population estimates, and the extent of that bias cannot
be quantified.

7.3 Sampling techniques

Random sampling means that each unit of interest within the population has the same probability
of selection into the sample as every other unit. The probability of selection of individual units
must not differ. This is irrespective of accessibility, ease of collection or other differences that may
exist between individuals. There are several important considerations to take into account before
collecting a random sample:

• The target population must be defined and identified.

• A study group (sometimes called a study population) that is representative of the target pop-
ulation must be identified. The study group should not differ in composition from the target
population.

• A sampling frame is produced. The sampling frame provides a means for identifying every unit
of interest (sampling unit) within the study group.

• Sampling units are selected from the sampling frame using a random (probabilistic) approach
such that each sampling unit within the sampling frame has an equal chance of being selected.

73
Methods of randomisation

There are two principal techniques for random sampling, physical randomisation and the use of
random numbers. Physical randomisation is a process where sampling units are selected using
physical systems that contain random elements. These include the selection of numbered marbles
from a bag, the use of a die, or the toss of a coin.
Random numbers are a sequence of numbers comprising individual digits with an equal chance that
any number from 0 to 9 will be present. Tables of random numbers can be used for sample selec-
tion. Some computer programs can generate random numbers. These programs use algorithms to
produce the sequence of numbers. The sequence of numbers that is generated depends upon the
value chosen as the starting value for the algorithm (the seed value). Whilst there is an equal prob-
ability that any digit from 0 to 9 will be present in a position chosen at random from the sequence,
the actual digit present at each point of the sequence is determined by the seed value. In other
words, the exact sequence of random numbers can be reproduced if the process is repeated using
the same seed value. Computer-generated random numbers are frequently called pseudo-random
numbers for this reason.

Replacement

Samples may be taken in one of two ways: with replacement or without replacement. In sampling
with replacement, each selected unit is examined and recorded and then returned to the sampling
frame. These units may then be selected into the sample again.
In sampling without replacement, each selected unit is examined and recorded and then withdrawn
from the sampling frame. These units are excluded from selection into the sample again. Intuitively,
sampling without replacement is the most logical — it is better to have different information from new
animals as opposed to having copies of information obtained from the repeated sampling of a single
animal. However, there are statistical reasons why sampling with replacement may be employed
in certain circumstances. These reasons relate to the mathematics of the estimation process. In
sampling with replacement the probability of selection of a unit remains the same from the first
selection through to the last selection. The distribution of results within the final sample is described
by the binomial distribution. In sampling without replacement, the probability of selection of the next
unit changes each time a selection is made. This is due to a reduction in size of the denominator as
each unit is drawn. The distribution of results is described by the (more complex) hypergeometric
distribution.
The difference between the two sampling procedures is not important when samples are drawn from
large populations. Often, the binomial distribution is used to approximate the hypergeometric distri-
bution when analysing the results of samples drawn without replacement from large populations.

Probability proportional to size sampling

When clusters differ widely with respect to the number of units that they contain, unequal probability
sampling of clusters will often result in estimates of population characteristics, especially popula-
tion totals, that have lower standard errors than those obtained from sampling clusters with equal
probability. Probability proportional to size sampling avoids these problems.

74
Say, for example, suppose you need to take a random sample of three herds from a list of the 10
herds shown Table 23. First divide the total population (6700) by the number of herds to be selected
(3) to obtain a sampling interval (6700 ÷ 3 = 2233). Next choose a random number between 1 and
2233. Suppose the chosen number is 1814. This should be fitted in position in the list to identify the
first herd in the sample. Since 1814 lies between 1601 and 1900, the first selected cluster is herd 4.
Next, add the sampling interval to the initial random number: 1814 + 2233 = 4047. The next cluster
to be selected is herd number 6. Add the sampling interval again: 4047 + 2233 = 6280 and herd 10
is chosen.
Table 23: A cumulative list of herd sizes.

Herd n Cumulative n
1 1000 1000
2 400 1400
3 200 1600
4 300 1900
5 1200 3100
6 1000 4100
7 1600 5700
8 200 5900
9 350 6250
10 450 6700

Note that when this technique is used it is possible for the same cluster to be selected twice if the
cluster has a population size that is greater than the sampling interval. This is unlikely to happen if
the proportion of clusters selected is small, unless one cluster is very much larger than the others.
If this occurs, you should select two subsamples of subjects from within the cluster. It is not valid to
select another cluster instead, or to repeat the sampling procedure until no clusters are repeated,
since either of these two approaches invalidates the required probabilities.
If no estimate of cluster sizes is available it will be impossible to carry out selection proportional
to size and clusters must be selected by simple random sampling methods. If this is the case,
responses will need to be weighted in any analyses that are undertaken. This requires a count of
the total number of sampling units in each selected cluster.

7.4 Sample size calculations

The choice of sample size involves both statistical and non-statistical considerations. Non-statistical
considerations include the availability of time, money, and resources. Statistical considerations in-
clude the required precision of the estimate, and the variance expected in the data. In descriptive
studies we need to specify the desired level of confidence that the estimate obtained from sampling
is close to the true population value (1 − α). In analytical studies we may also be interested in the
power (1 − β) of the study to detect real effects.

75
Simple and systematic random sampling

Formulae to derive sample sizes appropriate to estimate population parameters (population total,
mean, and proportion) on the basis of a simple random sample are as follows:

Total
z 2 SD2
n>
2

Mean
z 2 SD2
n>
2

Proportion
z 2 (1 − Py )Py
n>
2

z : the reliability coefficient (e.g. z = 1.96 for an alpha level of 0.05)

SD: the population standard deviation of the variable of interest
: the maximum absolute difference between the sample estimate and the unknown population value
Py : the unknown population proportion
We want to estimate the mean bodyweight of deer on a farm. We anticipate the standard deviation of body weight in
farmed deer of this age is around 30 kg. We would like to be 95% certain that our estimate is within 10 kg of the true
mean. How many deer should we include in our sample?
z = 1.96
SD = 30
 = 10
n = (z2 × SD2 ) / 2
n = (1.96 × 1.96 × 30 × 30) ÷ (10 × 10)
n = 34
We need to sample 34 deer.

We want to estimate the sero-prevalence of brucellosis in a population of cattle. The expected prevalence is 15% and we
would like to take enough samples to be 95% sure that our estimate is within 20% of the actual prevalence of disease.
How many cattle should be included in our sample?
z = 1.96
Py = 0.15
Absolute error = 0.20 × 0.15 = 0.03
n = [ z2 × (1 - Py) × Py ] / 2
n = [ 1.96 × 1.96 × (1 - 0.15) × 0.15 ] ÷ (0.03 × 0.03)
n = 544
We need to sample 544 cattle.

76
Detection of disease

Veterinarians are frequently asked to test groups of animals to confirm the absence of disease. The
number of animals that should be tested to provide a specified level of confidence that disease is
detected is given by:
1 D−1
n = (1 − α D ) × (N − ) (18)
2
Where:
N : the population size
α: 1 - confidence level (usually α = 0.05)
D: the estimated minimum number of diseased animals in the group (that is, population size × the
minimum expected prevalence)
What is the approximate number of animals that should be tested in a herd of 200 to be 95% confident that at least one
diseased animal will be found if the expected prevalence is 20%?
N = 200
α = 0.05
D = 0.20 × 200 = 40
n = (1 - α1/D) × (N - [D - 1] / 2)
n = (1 - 0.051/40) × (200 - [40 - 1] / 2)
n = 0.072 × 180.5
n = 13
A minimum of 13 animals need to be tested.

The above formula assumes that the test being used has perfect ability to detect an animal as
diseased, if it really is (that is, the test has a sensitivity of 1.0). The number to be tested can be
adjusted to account for an imperfect testing procedure, by multiplying the result from the ‘standard’
equation by the reciprocal of the test sensitivity. See Chapter 8
for details.
In the example above, we worked out that 13 animals need to be tested to be 95% confident that at least one disease
animal will be found if the expected prevalence of disease was 20%. How many animals should be tested if sensitivity
of the diagnostic test is 0.90?
n = 13
Se = 0.90
n’ = n × (1 / Se)
n’ = 13 × (1 / 0.90)
n’ = 15
Using a diagnostic test with a sensitivity of 0.90, a minimum of 15 animals need to be tested.

77
8 Diagnostic tests

By the end of this unit you should be able to:

• Explain what is meant by the terms sensitivity, specificity, and positive and negative predictive value as applied
to diagnostic tests.
• Given the appropriate formula and test results presented in a 2 × 2 table, calculate and interpret test sensitivity,
specificity, positive and negative predictive value.
• Explain the difference between apparent prevalence and true prevalence. Given data presented in a 2 × 2 table,
calculate apparent and true prevalence.
• Explain what is meant by parallel and series interpretation of diagnostic tests. Provide examples of where parallel
and series test interpretation is used (or would be useful).
• Explain how you would estimate the pre-test probability of the various diseases you might encounter in clinical
practice.
• Using a nomogram and an estimate of the pre-test probability of disease, determine the the post-test probability
of disease in an individual.

A test may be defined as any process or device designed to detect (or quantify) a sign, substance,
tissue change, or body response in an animal. Tests included:

• Routine examination of an animal or premises.

• Questions posed during history taking.

• Clinical signs.

• Laboratory findings — haematology, serology, biochemistry, histopathology.

• Post mortem findings.

If tests are to be used in a decision-making context, the selection of an appropriate test should be
based on its ability to alter your assessment of the probability that a disease does or does not exist.

8.1 Accuracy and precision

The accuracy of a test relates to its ability to give a true measure of the substance being measured.
To be accurate, a test need not always be close to the true value, but if repeat tests are run, the
average of the results should be close to the true value. An accurate test will not over- or under-
estimate the true value. Results from tests can be ‘corrected’ if the degree of inaccuracy can be
measured and the test results adjusted accordingly.
The precision of a test relates to how consistent the results of the test are. If a test always gives the
same value for a sample (regardless of whether or not it is the correct value), it is said to be precise.

78
Accuracy

Assessment of test accuracy involves running the test on samples with a known quantity of sub-
stance present. These can be field samples for which the quantity of substance present has been
determined by another, accepted reference procedure. Alternatively, the accuracy of a test can be
determined by testing samples to which a known quantity of a substance has been added. The
presence of background levels of substance in the original sample and the representativeness of
these ‘spiked’ samples make this approach less desirable for evaluating tests designed for routine
field use.

Precision

Variability among test results might be due to variability among results obtained from running the
same sample within the same laboratory (repeatability) or variability between laboratories (repro-
ducibility). Regardless of what is being measured, evaluation of test precision involves testing the
same sample multiple times within and/or among laboratories.

8.2 Diagnostic test evaluation

The two key requirements of a diagnostic test are: (1) the test will detect diseased animals correctly,
and (2) the test will detect non-diseased animals correctly. To work out how well a diagnostic test
performs, we need to compare it with a ‘gold standard.’ A gold standard is a test or procedure that
is absolutely accurate. It diagnoses all diseased animals that are tested and misdiagnoses none.
Histopathological and microbiological examination of the small intestine is generally regarded as the gold standard test
for Johne’s disease in cattle.
Histopathological examination of the brain stem is the gold standard test for bovine spongiform encephalopathy.

Once samples are tested using the gold standard and the test to be evaluated, a 2 × 2 table can be
constructed, allowing test performance to be quantified. The usual format is shown in Table 24.

Table 24: Diagnostic test data presented in a 2 × 2 table format.

Gold std + Gold std - Total

Test positive a b a+b
Test negative c d c+d
Total a+c b+d a+b+c+d

Understanding some of the terms used to describe the performance of diagnostic tests will be helped
if you think of test results and disease status in a population in the format shown in Figure 37.

Sensitivity

The sensitivity of a test is defined as the proportion of subjects with disease that test positive
[p(T + |D+ )]. A sensitive test will rarely misclassify animals with the disease. Sensitivity is a measure

79
Figure 37: There are 100 individuals in a population. Four individuals are diseased (black solid circles), 96 are healthy
(open circles). All 100 members of the population are tested: 10 return a positive test (pink [dark] shading); 90 return a
negative test (green [light] shading).

of accuracy for predicting events.

a
Sensitivity = (19)
(a + c)

Sensitivity is:

• The conditional probability of a positive test, given the presence of disease.

• The likelihood of a positive test in a diseased animal.

• The proportion of animals with disease that have a positive test for the disease.

• The true positive rate (relative to all animals with disease).

Specificity

The specificity of a test is defined as the proportion of subjects without disease that test negative
[p(T − |D− )]. A highly specific test will rarely misclassify animals that are not diseased.

d
Specificity = (20)
(b + d)

Specificity is:

• The conditional probability of a negative test, given the absence of disease.

• The likelihood of a negative test in an animal without disease.

• The proportion of animals without the disease that have a negative test for the disease.

80
 (a) Sensitivity (b) Specificity

Figure 38: Explanation of test sensitivity and specificity: (a) sensitivity is the proportion of disease-positive individuals who
test positive. In this example 3 out of 4 disease positive individuals test positive, so test sensitivity is 0.75; (b) specificity is
the proportion of disease-negative individuals who test negative. In this example 89 out of 96 disease negative individuals
test negative, so test specificity is 0.93. Key: black circles – diseased, open circles – healthy, pink [dark] shading – test
positive, green [light] shading – test negative.

• The true negative rate (relative to all animals without disease).

Sensitivity and specificity are inversely related and in the case of test results measured on a contin-
uous scale they can be varied by changing the cut-off value (Figure 39). In doing so, an increase
in sensitivity will often result in a decrease in specificity, and vice versa. The optimum cut-off level
depends on the diagnostic strategy. If the primary objective is to find diseased animals (that is, to
minimise the number of false negatives and accept a limited number of false positives) a test with
a high sensitivity is required. If the objective is to make sure that every test positive is ‘truly’ dis-
eased (minimise the number of false positives and accept a limited number of false negatives) the
diagnostic test should have a high specificity.

Positive predictive value

The positive predictive value is the proportion of subjects with positive test results which have the
disease.
a
Positive predictive value = (21)
(a + b)

Positive predictive value is:

• The predictive value of a positive test.

• The post-test probability of disease following a positive test.

• The posterior probability of disease following a positive test.

81
Figure 39: Test results measured on a continuous scale, showing the distribution of results that might be obtained for
healthy and diseased individuals. The cut-off value for the test is shown by the vertical solid line: those individuals with
a result less than the cut-off value are diagnosed as non-diseased, those individuals with a result greater than the cut-off
value are diagnosed as diseased. Using this diagnostic test, disease-negative individuals with a test result greater than
the cut-off value (‘A’ in the left-hand plot) will be false positive. Disease-positive individuals with a test result less than the
cut-off value (‘B’ in the right-hand plot) will be false negatives.

Negative predictive value

The negative predictive value is the proportion of subjects with negative test results which do not
have the disease.
d
Negative predictive values = (22)
(c + d)

Negative predictive value is:

• The predictive value of a negative test.

• The post-test probability of no disease following a negative test.

• The posterior probability of no disease following a negative test.

Predictive values quantify the probability that a test result for a particular animal correctly identifies
the condition of interest. Estimation of predictive values requires knowledge of sensitivity, specificity
and the prevalence of the disease of interest in the population. The effect of prevalence on predictive
values is considerable. Given a prevalence of disease in a population of around 30% and we are
using a test with 0.95 sensitivity and 0.90 specificity, the predictive value of a positive test would be
0.80 and the predictive value of a negative test would be 0.98. If the prevalence of disease is only
3% and the test characteristics remain the same, the predictive value of a positive and negative test
will be 0.23 and 0.99, respectively.

82
 (a) Positive predictive value (b) Negative predictive value

Figure 40: Explanation of positive and negative predictive value: (a) positive predictive value is the proportion of test-
positive individuals who are actually disease positive. In this example 3 out of 10 test positive individuals are diseased, so
the positive predictive value of the test is 0.30; (b) negative predictive value is the proportion of test-negative individuals
who are actually disease negative. In this example 89 out of 90 test negative individuals are actually disease negative, so
the negative predictive value of the test is 0.99. Key: black circles – diseased, open circles – healthy, pink [dark] shading
– test positive, green [light] shading – test negative.

Sensitivity and specificity are properties of a test. Sensitivity and specficity don’t change with prevalence.
If the prevalence increases, positive predictive value increases and negative predictive value decreases. If the preva-
lence decreases, positive predictive value decreases and negative predictive value increases.
The more sensitive a test, the better its negative predictive value. The more specific a test, the better its positive
predictive value.

8.3 Prevalence estimation

The estimate of disease prevalence determined on the basis of an imperfect test is called the ap-
parent prevalence. Apparent prevalence is the proportion of all animals that give a positive test
result. It can be more than, less than, or equal to the actual proportion of diseased animals, the
true prevalence. If sensitivity and specificity of a test are known, true prevalence can be calculated
using the Rogan and Gladen (1978) formula:

AP − (1 − Sp) AP + Sp − 1
p(D+ ) = = (23)
1 − [(1 − Sp) + (1 − Se)] Se + Sp − 1

Where:
AP : apparent prevalence
Se: sensitivity (0 - 1)
Sp: specificity (0 - 1)

83
Figure 41: Relationship between prevalence and positive predictive value for tests of different sensitivities and specifici-
ties.

Individual cow somatic cell counts (ICSCC) are used as a screening test for subclinical mastitis in dairy cattle. This test
has a sensitivity of 0.90 and a specificity of 0.80. The apparent prevalence of mastitis in this herd using the screening
test is 23 cases per 100 cows. True prevalence p(D+) may be calculated as follows:
AP = 0.23
Se = 0.90
Sp = 0.80
p(D+) = (AP + Sp - 1) ÷ (Se + Sp - 1)
p(D+) = (0.23 + 0.80 - 1) ÷ (0.90 + 0.80 - 1)
p(D+) = 0.03 ÷ 0.70
p(D+) = 0.04
The true prevalence of mastitis in this herd is 4 cases per 100 cows.

The Rogen-Gladen approach described above doesn’t perform well when you’re dealing with a dis-
ease with a very low prevalence (say 1 case per 100 animals at risk or less). In these situations it’s
necessary to use a Bayesian approach to estimate true prevalence.

8.4 Diagnostic strategies

Clinicians commonly perform multiple tests to increase their confidence that a patient has a particular
diagnosis. When multiple tests are performed and all are positive, the interpretation is straightfor-
ward: the probability of disease being present is relatively high. It is far more likely however, that
some of the tests return a positive result and others will be negative. We can deal with this problem
by interpreting test results in parallel or series.

Parallel interpretation

Parallel interpretation means that when multiple tests are run an individual is declared positive if at
least one of the multiple tests returns a positive result. Interpreting test results in parallel increases

84
the sensitivity and therefore the negative predictive value for a given disease prevalence. However,
specificity and positive predictive value are lowered. As a consequence, if a large number of tests
are performed and interpreted in this way then virtually every individual will be considered positive.

Serial interpretation

Series interpretation means that when multiple tests are run an individual is declared positive if
all tests return a positive result. Series interpretation maximises specificity and positive predictive
value which means that more confidence can be attributed to positive results. It reduces sensitivity
and negative predictive value, and therefore it becomes more likely that diseased animals are being
missed.

Figure 6: ROC curve (left) and test sensitivity and specificity plotted against various possible
cutpoints (right). From the right-hand graph, a cutpoint of 0.40 yields the optimum test
sensitivity and specificity.:

8.5 Screening versus diagnosis

In clinical practice, tests tend to be used in two ways. Screening tests are those applied to appar-
ently healthy members of a population to detect the presence of disease, disease-causing agents, or
subclinical disease. Usually, those animals that return a positive to such tests are subject to further
in-depth diagnostic work-up. Diagnostic tests are used to confirm or classify disease status, provide
a guide to selection of treatment, or provide a prognosis. In this setting, all animals are ‘abnormal’
and the challenge is to make a correct diagnosis.
With a screening and confirmatory test strategy a test is applied to every animal in the population to
screen the population for positives. Ideally, this test should be easy to apply and low in cost. It also
should be a highly sensitive test so that it misses only a small number of diseased animals. Its speci-
ficity should still be reasonable, so that the number of false positives subjected to the confirmatory
test remains economically justifiable.
Individuals that return a negative result to the screening test are regarded as true negatives and
not submitted to further examination. Animals positive to the screening test are subjected to a con-
firmatory test. The confirmatory test can require more technical expertise and more sophisticated
equipment, and be more expensive, because it is only applied to a reduced number of samples.
But it has to be highly specific, so that any positive reaction to the confirmatory test is considered a
definitive positive.
During the early phase of disease control programs (e.g. programs to eradicate tuberculosis) the
apparent prevalence will be higher than the true prevalence, as a consequence of test specificity
being less than 1.00. As the program continues, test positive animals are identified and culled which
results in a decrease in true prevalence. As true prevalence declines the positive predictive value
of testing declines, increasing the proportion of false positives. At this stage of the control program
a highly specific test is required. In some cases it may be necessary to use a number of tests
interpreted in series to increase specificity.

85
SPINS and SNOUTS. SPecific tests are needed to rule a diagnosis IN, and highly SeNsitive tests are needed to rule
them OUT. When a disease is rare however (with a prevalence of less than 0.01) the specificity of a test is rarely high
enough to give adequate positive predictive value. Only the sensitivity is useful in the rare disease case. To remember
this:

Thinking about,
SPIN and SNOUT
In cases where
Disease is rare.
Don’t use SPIN,
But keep SNOUT in.

Positive and negative predictive values are more useful to the clinician than sensitivity and specificity. Predictive values
vary with prevalence, a common mistake is to assume they are fixed. You need to know the prevalence of disease to
derive valid estimates of positive and negative predictive value.

8.6 Likelihood ratios

Diagnostic testing is often undertaken to help us decide whether or not an individual is diseased.
Because diagnostic tests are imperfect (that is, false positives and false negatives occur) we should
move away from the ‘test positive = disease positive’, ‘test negative = disease negative’ paradigm
and think about testing as a process that provides us with a probability estimate of the presence of
disease in an individual. Likelihood ratios provide a means for doing this.
The likelihood ratio for a positive test tells us how likely we are to find a positive test result in a
diseased individual compared with a positive test result in a non-diseased individual. The likelihood
ratio for a positive test equals (sensitivity) divided by (1 - specificity). The likelihood ratio for a
negative test tells us how likely we are to find a negative test result in a diseased individual compared
with a negative test result in a non-diseased individual. The likelihood ratio for a negative test equals
(1 - sensitivity) divided by the specificity. Thus:
Se
LR+ = (24)
1 − Sp
1 − Se
LR− = (25)
Sp
Where:
Se: sensitivity (0 - 1)
Sp: specificity (0 - 1)
Likelihood ratios (LR) can be calculated using single cut-off values, so we have one pair of likelihood
ratios: one for a positive (LR+) and another for a negative test result (LR-). More information can
be extracted from the diagnostic test by using multilevel likelihood ratios. In this case ranges of test
results will have associated likelihood ratio values.
Likelihood ratios can be used to apply Bayes’ Theorem to the interpretation of diagnostic test results.
A brief outline of Bayes’ Theorem in this context is as follows. To increase your confidence in an
individual’s disease status you use a diagnostic test. The first step is that you nominate the pre-test
probability that the individual has the disease. Second, you apply the diagnostic test, which has
certain characteristics (i.e. a positive and negative likelihood ratio). The third and final step is to

86
use the pre-test probability and the likelihood ratio to derive a post-test probability of disease in the
individual, given the result of testing.
To do these calculations by hand we need to convert the pre-test probability of disease into a pre-test
odds, multiply the pre-test odds by the appropriate likelihood ratio which then gives you a post-test
odds. Finally, the post-test odds is converted to a post-test probability. The relationship between
odds and probability is as follows:

Probability of event
Odds of event = (26)
1 - Probability of event

Odds of event
Probability of event = (27)
1 + Odds of event

Individual cow somatic cell counts (ICSCC) are used as a screening test for sub-clinical mastitis in dairy herds. A client
has a herd of dairy cows where the (true) prevalence of subclinical mastitis is estimated to be around 5%. Your herd
testing authority tells you that the positive likelihood ratio for a cell count of 300 – 400 cells/mL is 14.50.
You are called to examine an individual cow from this herd and find that she has an ICSCC of 320,000 cells/mL. What is
the probability that this cow really has mastitis? The pos-test probability of mastitis in this cow is determined as follows:

1. The pre-test probability of mastitis: 50 ÷ 1000 = 0.05.

2. The pre-test odds of mastitis: 0.05 ÷ (1 - 0.05) = 0.053.
3. The post-test odds of mastitis given a positive test result: pre-test odds × LR(+) = 0.053 × 14.5 = 0.76.
4. The post-test probability of mastitis given a positive test result: 0.76 ÷ (1 + 0.76) = 0.43.

The post-test probability of a cow with a ICSCC of 320,000 cells/mL being mastitic is around 43%.

A nomogram (Figure 42) provides a convenient short cut for calculating post-test disease probabil-
ities. The nomogram is a chart comprised of three scaled parallel lines. On the far left hand line
you mark the position that corresponds to the pre-test probability of disease. On the central line you
mark the positive (or negative) likelihood ratio of the test you’re using. If you draw a straight line from
the pre-test probability estimate through the position of the likelihood ratio on the central line you
can then read off the post-test probability of disease from the far right hand line.
A nice feature of this approach to evaluating diagnostic test information is that sequential testing can
be easily handled. In this situation the post-test probability of disease from the first round of testing
becomes the pre-test probability for the second round of testing.
To continue the mastitis example described above lets imagine that we examine our cow and as part of that examination
we test milk from each quarter using a rapid mastitis test (RMT). We are told that the sensitivity and specificity of the
RMT is 0.70 and 0.80, respectively. Our cow returns a positive result to the RMT. What is the probability of this cow
being mastitic, given this additional information?
The likelihood ratio of a positive RMT is (0.70 / 1 - 0.80) = 3.5. If the pre-test probability of disease is 43% we can use
the nomogram to estimate the posterior probability of disease, given a positive test, as 72%. We’re now 72% certain
that this cow has mastitis.

The advantage of the likelihood ratio method of test interpretation is that we can better appreciate
the value (i.e. the increase in post-test probability) provided by each diagnostic test that is applied (in
the above example, ICSCC provided more diagnostic information than the RMT). If the cost of each
test applied is known the cost per unit increase in post-test probability can be determined, enabling
us to be more objective in our use of diagnostic resources.

87
 Figure 42: Nomogram for post-test probability calculations using likelihood ratios of a positive test result.

Figure 43: Diagram showing how the estimated probability of disease changes after applying a series of diagnostic tests.
In our example of the cow with mastitis, we had a prior belief that the probability of the cow being mastitic was 5%. After
considering the ICSCC result this probability increased to 43%. After applying a rapid mastitis test and getting a positive
result, the probability of the cow having mastitis increased to 72%.

88
9 Outbreak investigation methods

By the end of this unit you should be able to:

• Describe the steps to take during an outbreak investigation, including description of the outbreak by animal, place
and time.
• Explain why it is important to establish a case definition when investigating a disease outbreak.
• List methods you might use to enhance surveillance once an outbreak of disease has been identified.

An outbreak is a series of disease events clustered in time. During an outbreak the investigator asks
the questions:

• What is the problem?

• Can something be done to control it?

• Can future occurrences be prevented?

These notes outline an approach to investigating outbreaks of disease in animal populations. Al-
though the term outbreak implies a sudden (and possibly spectacular) event (e.g. an outbreak of
botulism in feedlot cattle), be aware that outbreaks can be of a more insidious nature: some caus-
ing subclinical losses in a population of animals over an extended period before being identified,
characterised and investigated.

9.1 Verify the outbreak

What is the illness?

Once a suspected outbreak is identified, identifying the specific nature of the illness is an important
early step. An attempt should be made to characterise cases (leading towards a formal case defini-
tion, see below). It may not be possible to make a definitive diagnosis at this stage. What is required
is a working definition of the disease or syndrome: for example ‘ill thrift in recently weaned calves’
or ‘sudden death in grower pigs.’

Is there a true excess of disease?

The first issue to be certain of is whether or not the outbreak is genuinely an unusual event worthy
of special attention. The number of cases per unit time should be substantially greater than what
is normal for the group of individuals under investigation. It is common to have owners and others
concerned about a possible outbreak which is transient increase in the normal level of endemic
disease.

89
9.2 Investigating an outbreak

Establish a case definition

A case definition is the operational definition of a disease for study purposes. A good case defini-
tion has two parts: (1) it specifies the characteristics of the population at risk, and (2) it specifies
what distinguishes cases from other members of the population. A case definition ensures that the
outcome of interest is consistently defined across space (e.g. among different investigation centres
in a large scale outbreak) and over time. Consistency in case definition is important since it will
allow the incidence of disease to be measured which in turn allows responses to control efforts to
be monitored.
In an outbreak of this severe and often fatal pneumonia in delegates attending the 58th annual meeting of the American
Legion, Department of Pennsylvania a case was considered Legionnaires’ disease if it met clinical and epidemiologic
criteria. The clinical criteria required that a person have onset between 1 July and 18 August 1976, an illness charac-
terised by cough and fever (temperature of 38.9◦ C or higher) or any fever and chest x-ray evidence of pneumonia. To
meet the epidemiologic criteria, a patient either had to have attended the American Legion Convention held 21 – 24 July
1976, in Philadelphia, or had to have entered Hotel A between 1 July 1976 and the onset of illness.
Reference: Fraser DW, Tsai TR, Orenstein W, Parkin WE, Beecham HJ, Sharrar RG, Harris J, Mallison GF, Martin SM,
McDade JE, Shepard CC, Brachman PS (1977). Legionnaires’ disease — description of an epidemic of pneumonia.
New England Journal of Medicine, 297:1189-1197.

The following case definitions for infection with swine influenza A (H1N1) virus were defined by the New Zealand Ministry
of Health on 22 May 2009:
A confirmed case of swine influenza A (H1N1) virus infection is defined as a person with an acute respiratory illness with
reference laboratory confirmed swine influenza A (H1N1) virus infection by one or more of the following tests: real-time
PCR, viral culture, a four-fold rise in swine influenza A (H1N1) virus specific neutralising antibodies.
A probable case of swine influenza A virus infection is defined as a person who meets the suspected case definition and
tests positive for influenza A.
A suspected case of swine influenza A virus infection is defined as a person with an acute respiratory illness who either:
(a) has developed symptoms within 7 days of travel to an area where there are confirmed cases and confirmed or
suspected local transmission of swine influenza A; or (b) has an acute respiratory illness who is considered to be a close
contact of a probable or confirmed case.

Enhance surveillance

When it is suspected that an outbreak is occurring, enhanced surveillance can be useful to iden-
tify additional cases. Enhanced surveillance may involve both heightening awareness to increase
passive case reports and implementing targeted surveillance. Techniques include directly contact-
ing field practitioners by telephone, facsimile or email, via health department web pages and email
discussion groups. For large outbreaks media releases (print, television, radio) can be extremely
effective.

Describe outbreak according to individual, place and time

Record the physical layout of the farm premises. Draw and label a map of all pens, pastures or
other physical characteristics that demarcate groups of animals. Identify the pens or areas using the
producer’s system (otherwise the results from groups may be meaningless to all parties). If feeds

90
are involved, indicate storage locations. Record the current numbers of animals in each pen or
paddock and the intended maximum capacity of these areas. These numbers are useful for defining
denominators when you want to calculate incidence or prevalence.
Record the animal ‘calendar’ intended by management for the relevant animals. The animal calendar
is what happens to animals as they move through their production cycle. These are the what’s and
when’s (what age or days in production cycle) for being moved, fed, vaccinated or otherwise handled
(e.g. dried off, dehorned, bred, examined for pregnancy) and otherwise exposed to potential risk
factors. The easiest way to do this is to start at a point in the production cycle and work your way
around the cycle, recording the management policy for each event. Double-check later with the
people who actually carry out these procedure to see if this is indeed what happens to them.
Record the policy that determines when or why animals are routinely moved from one group to
another during normal operation. For example, the policy for drying cows off may be that lactating
cows are dried off when they fall below a particular production level or when they reach so many
days before calving (whichever comes first). A related issue is how often and on what schedule
is production assessed and cows moved. For example, is this done on a certain day of the week,
monthly after testing or when? If investigating a calf scour problem, the policy for moving cows from
the high string to the low string is not relevant but the policy for moving cows during the dry period
may be.
For those policies that may be related to the problem at hand compare evidence of the actual practice
to the management policy. Has this policy been enforced well, is enforcement variable, or is the
policy becoming lax? Continuing the example above, determine if the dry off policy is actually
practiced by assessing the lengths of dry periods of recently calved cows. Management often has
one policy (e.g. calves are weaned at 30 days) but employees may be executing another.
Obtain specific information on practices. For example, if the transmission of an infectious agent by
treatment equipment is potentially involved, ask how and when the equipment is sterilized. Between
animals? Is it washed with soap first? What concentration of what disinfectant is used? How long is
the contact time?
Collect historical, clinical and productivity data on those individuals that are affected (cases) and
those that are not affected (non-cases). If possible, all cases of diseased animals should be in-
cluded in the investigation. If there are large numbers of unaffected individuals you may select a
representative sample of unaffected individuals for examination (controls). You may consider match-
ing controls with some characteristic of the cases e.g. age and gender.
Plot an epidemic curve by identifying the first detected case (index case) and then graphing subse-
quent numbers of cases per day or per week from the index case through to the end of the outbreak.
Does the epidemic show common source or propagated properties?

Develop hypotheses about the nature of exposure

At this stage, you will probably have some suspicions about what has caused the outbreak — that
is, you will have started to form some hypotheses. Your next job is to test these hypotheses using
the various analytical techniques described below.

91
Conduct analytical studies

Part of the data collection procedure above will have entailed collecting individual-level details such
as age, sex, breed, date of parturition, stage of production. Individuals should be categorised ac-
cording to the presence of each attribute. Attack rate tables divide the cohort of interest into exposed
and non-exposed groups. Attack rates are then calculated for each exposure by dividing the number
diseased by the group size (Table 25).

Table 25: Attack rate table for an outbreak of food poisoning.

Food Exposed Unexposed

Ill Well Total AR (%) Ill Well Total AR (%)
Ham 36 5 41 88 2 11 13 15
Salad 40 4 44 91 9 6 15 60
Prawns 16 15 31 52 10 13 23 43

The exposure which is most likely to have served as a vehicle for an outbreak is that with the greatest
difference in attack rate for exposed and unexposed individuals. In Table 25 the greatest difference in
attack rate is for the ham-exposed and ham-unexposed groups. This would support the hypothesis
that ham was the source of this outbreak of food poisoning. An alternative is to calculate the risk
ratio of disease for each exposure. Essentially this is the attack rate for the exposed individuals
divided by the attack rate for unexposed individuals — the exposure with the highest risk ratio being
the likely vehicle for the outbreak. It is also useful to calculate the population attributable fraction
for each exposure. This will identify the percent of the risk of disease in the exposed group that is
due to exposure. The closer this value is to 100% the more likely the exposure accounted for the
outbreak.

9.3 Implement disease control interventions

At this stage it may be possible to produce a hypothesis regarding the cause of the outbreak and to
implement controls on the basis of these hypotheses. Provide written and verbal instructions to your
client detailing your approach to controlling the outbreak. Gardner (1990b) provides a nice summary
of how to write up an outbreak investigation.
Ensure the appropriate measures are being taken to monitor the response to your interventions. This
allows you to monitor how things are going and to revise your control plan at the first opportunity if
things don’t work out as expected. If further investigation is warranted then other epidemiological
studies (case-control, prospective cohort etc) may be designed and implemented. You may also use
more complex analytical techniques to analyse data that has already been collected (multivariate
techniques).

92
Figure 44: Report of an outbreak of Salmonellosis in humans arising from a contaminated buffet lunch. Source: The
Globe and Mail (Toronto, Canada) Thursday 19 May 2005.

93
10 Critical appraisal

By the end of this unit you should be able to:

• Describe, in your own words, the four main areas that should be considered when appraising the scientific
literature.
• Explain what is meant by the terms internal and external validity.
• Explain the difference between the eligible population and the study population.

‘The sin is not in doing the research nor even in publishing the results. The sin is
in believing your results.’ S ANDER G REENLAND

Reading the literature is necessary to keep up to date with new developments and to learn more
about a particular area of science that interests us. Fortunately, there appears to be no shortage of
literature available to read, and our ability to source this literature easily has been facilitated by the
Internet (either in the form of peer-reviewed articles published on-line by established journals or as
pre-print publications published by individuals on their own web pages). Although the Internet allows
information to be widely disseminated, the quality of that information varies widely. As a result, as
good scientists, we need to be discerning about what we read and (more importantly) what we
believe. A systematic method of appraising (or evaluating) the literature helps us to do this. These
notes outline a systematic approach to appraising the epidemiological literature, which consists of:

1. Describing the evidence;

2. Assessing the internal validity of the study;

3. Assessing the external validity of the study; and

4. Comparing the results with other available evidence.

These notes outline an approach for critically appraising epidemiological studies (i.e. those that
investigate the relationship between a set of exposures and a defined outcome). An excellent series
of articles providing guidelines for appraising other types of articles appeared in the British Medical
Journal in 1997 (Greenhalgh 1997a, 1997b, 1997c, 1997d, 1997e, 1997f, 1997g, and Greenhalgh
and Taylor 1997). Much of the technical material from these articles has been compiled into a very
readable textbook on the subject by the same author (Greenhalgh 2006).

10.1 Description of the evidence

The first step in evaluating a scientific article is to understand exactly what relationship was being
evaluated and what hypothesis was being tested. It should be relatively easy for the reader to identify
the exposure variable(s), the outcome variable and the study design (survey, case-control, cohort
study, clinical trial). You should be able to clearly define the following populations:

1. The study population: individuals who actually took part in the study.

94
 2. The eligible population: individuals who met the criteria to be included in the study.

3. The source population: the population from which eligible study subjects were drawn.

4. The external population: individuals not part of the source population (e.g. those in another
region or country) but who, based on the results of the study, you want to make generalisations
about.

The subjects that were studied in terms of source populations, eligibility criteria, and participation
rates of the different groups should be clearly stated.
Having defined the topic of study and those who took part, it is then useful to summarise the main
result — what is the result in terms of the association between exposure and outcome? It should
be possible to express the main result in a simple table and obtain from the paper the means to
calculate the appropriate measure of association (risk ratio, odds ratio, difference in proportions)
and the appropriate test of statistical significance.

10.2 Internal validity

Non-causal explanations

Once you know what the study is about the next step is to assess its internal validity — that is, are
the results valid for the subjects who were studied? Consider three non-causal mechanisms that
might have influenced the internal validity: bias, confounding and chance. The order of these
non-causal explanations is important. If there is severe observation bias, no analytical manipulation
of the data will overcome the problem (the study is fundamentally flawed). If there is confounding,
then appropriate analysis will (in most cases) overcome the problem. The assessment of chance
variation should be made on the main result of the study, after ruling out issues around bias and
confounding.

Causal explanations

Once the non-causal explanations for the results (bias, confounding, and chance) have been ruled
out, attention turns to considering the features of the study that support a claim that there is a causal
relationship between the exposure and outcome. Five causal mechanisms should be considered:

1. Is there a correct temporal relationship? For a relationship to be causal, the putative expo-
sure must act before the outcome occurs. In a prospective study design where exposed and
non-exposed subjects are compared, this requirement is established by ensuring that subjects
do not already have the outcome of interest when the study starts. The ability to clarify time
relationships is weaker in retrospective studies, and care is required to ensure that possible
causal factors did in fact occur before the outcome of interest. A difficulty in all study designs,
but more so in retrospective studies, is that the occurrence in biological terms of the outcome
of interest may precede the recognition and documentation of that outcome by a long and
variable period of time (e.g. some cancers).

95
 2. Is the relationship strong? A stronger association, that is a larger risk ratio, is more likely
to reflect a causal relationship. As a measured factor gets closer to a biological event on
the causal pathway, risk ratios become larger. The fact that a relationship is strong does not
always mean that the exposure-outcome relationship is causal, however if there is bias it must
be large and therefore easy to identify. If a strong relationship is due to confounding, either
the association of the exposure with the confounder must be very close, or the association of
the confounder with the outcome must be very strong.

3. Is there a dose-response relationship? In some circumstances the demonstration of a

smooth dose-response relationship may be a strong argument against an identified relation-
ship arising as a result of bias. In general, we should expect uni-directional dose-effect rela-
tionships and evidence that this is not the case should be considered carefully.

4. Consistency? A causal relationship will be expected to apply across a wide range of study
subjects. An association identified in one study that is consistent with the same associa-
tion identified in a different groups of subjects is supportive of causation. The difficulty with
consistency is that very large data sets are required to assess the similarity or otherwise of
associations in different subgroups of subjects. Even with adequate numbers, the subgroups
to be compared need to be defined on a priori grounds.

5. Specificity? It has been argued that specificity (that is, a given exposure produces a specified
outcome) provides good evidence for causality. Specificity may be useful, if we do not make
it an absolute criterion, as one exposure may produce various outcomes, and one outcome
may result from various exposures. The concept is often useful in study design: as a check
on response bias we may deliberately collect information on factors which we expect to be the
same in groups that we are comparing. Similar results across groups will indicate a lack of
observation bias.

10.3 External validity

With external validity we consider how appropriate it is to apply the results to populations apart from
the study population.

Can the results be applied to the eligible population?

The relationship between the study population and those who met the study inclusion criteria but did
not take part should be well documented. Losses due to non-participation have to be considered
carefully as they are likely to be non-random and the reasons for loss may be related to the exposure
and/or the outcome.

Can the results be applied to the source and external populations?

The important issue is not whether the individuals who were studied are ‘typical’ or representative
of the source population, but whether the association between outcome and exposure given by the
study participants is likely to apply to other groups. In general, the difficulties of applying results

96
from one group of subjects to another will be minimal for issues of basic physiology and maximal for
effects in which cultural and psycho-social aspects are dominant.

10.4 Comparison with other evidence

For many clinical questions a large amount of evidence is available which comes from different types
of studies. In these circumstances it is useful to consider a hierarchy of evidence. Given that studies
are adequately performed within the limitations of the design used, the reliability of the information
from them can be ranked (highest to lowest) as follows:

1. Randomised clinical trials.

2. Cohort and case-control studies.

3. Other comparative studies.

4. Case series, descriptive studies, clinical experience.

Randomised clinical trials, if properly performed on adequate numbers of subjects, provide the
strongest evidence of causation because of the unique advantages they provide in terms of over-
coming problems of bias and confounding.

Consistency

This is the most important characteristic used in the judgement that an association is causal. To say
that the study results are consistent requires that the association has been observed in a number
of different studies, each of which individually can be interpreted as showing a causal explanation.
Variation in study methodology and study populations make it unlikely that the same biases or con-
founding factors would be present in all of them. Lack of consistency argues against causality.

Plausibility

Plausibility refers to the observed association being biologically understandable on the basis of cur-
rent knowledge concerning its likely mechanisms. Be aware that any dramatically new observation
may be in advance of current biological thinking and its lack of plausibility may reflect deficiencies in
biological knowledge rather than error in observation. For example:

• John Snow effectively prevented cholera in London 25 years before the isolation of the cholera
bacillus and the general acceptance of the principle that the disease could be spread by water.

• Percival Pott demonstrated the causal relationship between exposure to soot and scrotal can-
cer some 150 years before the relevant carcinogen was isolated.

97
Coherency

An association is regarded as coherent if it fits the general features of the distribution of both the
exposure and the outcome under assessment. If lung cancer is due to smoking, the frequency of
lung cancer in different populations and in different time periods should relate to the frequency of
smoking in those populations at earlier relevant time periods.
If the exposure variable under study causes only a small proportion of the total disease, the over-
whelming influence of other factors may make the overall pattern inconsistent.

98
11 Exercise: outbreak investigation

This exercise has been adapted from Gardner (1990b).

A veterinarian in a mixed practice has been investigating an ongoing diarrhoea problem in neonatal
pigs in a 150-sow breeding/finishing herd. In the 12 months prior to the outbreak, 7% of litters had
diarrhoea but over recent weeks the proportion of litters affected has increased to about 40%. As
part of the investigation the veterinarian submitted 3 acutely affected pigs to the regional diagnostic
laboratory. Of the 3 pigs, 1 was infected with E. coli serotype 08 but other pathogenic bacteria and
viruses were not isolated from the other 2 pigs. Lesions in all 3 pigs were consistent with an acute
enteritis. The veterinarian asks you to assist.
As background to the problem, the veterinarian provides you with a map showing the layout of the
sheds, a description of normal management procedures, and recent records for farrowing sows as
detailed below:

11.1 The problem

Shed design. The shed has 16 concrete-floored pens (oriented in a single row in a west - east
direction. Pen 1 is near the entrance door at the western end of the shed and pens run in numerical
sequence to pen 16 which is located near the extraction fans. The pit underneath the sows is flushed
at least twice daily. During the study, pen 14 was under repair and was not used.
Management - treatments. Sows are moved into cleaned and disinfected pens in the farrowing shed
on about day 110 of gestation. Sows farrow with minimal supervision. On the first day of life, pigs
have their needle teeth clipped and are provided with heat lamps. No vaccines are given to sows or
baby pigs for control of enteric disease. Sows are fed ad libitum during lactation with a high energy
ration (15.5 MJ DE/kg). During gestation, they are fed about 2.0 to 2.5 kg of a lower energy ration
plus about 0.5 kg/day of recycled manure for control of enteric infections and parvovirus. Piglets
in litters with diarrhoea are treated with oral furazolidone and electrolytes are offered ad libitum in
shallow bowls in each pen.
Records. Records are provided from a recent set of 26 farrowings (April 2002) for you to examine
before your visit. Before April 2002 the records of diarrhoea were insufficiently detailed to be of value
in the current investigation.

11.2 Diagnosis

How valid are owner-diagnoses of scours-related deaths? How could you improve their validity in
the future?

11.3 Measures of disease frequency

Estimate the following rates from the data:

• The scours-specific mortality rate.

99
 • The proportional mortality rate for scours.
• The case fatality rate for scours.
• The proportion of litters affected with scours.
• The preweaning mortality rate.

11.4 Investigation

Outline your approach to investigating this diarrhoea problem (at this stage there is no need to
calculate any factor-specific rates). What initial conclusions or hypotheses did you formulate after
examining the history and laboratory findings, and temporal and spatial patterns of disease?

11.5 Measures of association

Analyse the records from the 26 April farrowings and calculate some factor-specific rates or relative
risks either by hand or by using computer software available for that purpose. For example:

• What was the risk ratio of scours in parity 1 litters, compared with litters from all other parities?

• What was the risk ratio of scours in litters from sick sows, compared with litters from healthy
sows?

• What was risk ratio of scours in large litters, compared with small litters?

• What was the risk ratio of scours in litters born in pens 1 – 8, compared with litters born in
pens 9 – 16?

Test the statistical significance of the difference between the two rates in each case. How helpful
are the data in allowing you to formulate better hypotheses? Could confounding be a problem and
how would you deal with it at this stage of the study?
We are interested in testing the hypothesis that the proportion of exposed individuals that are disease positive differs from
the proportion of non-exposed individuals that are disease positive. Because this is nominal (count) data, a chi-squared
test is the appropriate method to test this hypothesis. This involves three steps:
1. A statement of the null hypothesis: ‘The proportion of exposed individuals that are diseased does not differ from the
proportion of non-exposed individuals that are diseased.’
2. Calculation of a chi-squared test statistic. Using the standard notation the formula for the chi-squared test statistic for
data presented in a 2 × 2 table is:

n(ad − bc)2
χ21 = (28)
(a + c)(b + d)(a + b)(c + d)
3. We will use an alpha level of 0.05 to test this hypothesis and apply a one-tailed test. Specifying an alpha level of
0.05 means that there is a 5% probability of incorrectly rejecting the null hypothesis (when it is in fact true). The critical
value that separates the upper 5% of the χ2 distribution with 1 degree of freedom from the remaining 95% is 3.841 (from
statistical tables). Thus, if our calculated chi-squared test statistic is greater than 3.841 we can reject the null hypothesis
and accept the alternative hypothesis, concluding that the proportions diseased among exposed and non-exposed
individuals differ.

100
11.6 Recommendations

What recommendations, if any, would you make to your colleague and to his client based on your
findings (without the data from the clinical trial or cohort study)?

11.7 Clinical trial

Design either a clinical trial or a prospective cohort study to test one of your hypotheses in detail.

11.8 Financial impact

Estimate the financial impact of the losses due to diarrhoea in this set of 26 litters. The following
data has been provided:

101
 Litter Pen Sow Parity Farrow Born Weaned Death due to
Overlay Scours Other
1 9 124 1 03 Apr 02 12 9 1 2 0
2 4 121 1 03 Apr 02 9 6 1 2 0
3 12 76 3 04 Apr 02 8 8 0 0 0
4 13 164 2 05 Apr 02 11 9 0 2 0
5 16 27 6 06 Apr 02 7 7 0 0 0
6 1 18 4 09 Apr 02 10 6 0 4 0
7 a 7 3 2 10 Apr 02 14 8 2 2 2
8 3 69 8 10 Apr 02 10 9 1 0 0
9 11 13 5 11 Apr 02 8 8 0 0 0
10 2 101 3 12 Apr 02 12 7 2 1 2
11 8 83 6 14 Apr 02 11 10 1 0 0
12 5 79 2 15 Apr 02 11 11 0 0 0
13 10 62 4 18 Apr 02 9 8 1 0 0
14 a 6 74 1 18 Apr 02 10 7 0 3 0
15 4 27 1 19 Apr 02 9 6 0 3 0
16 15 61 7 23 Apr 02 6 5 1 0 0
17 12 52 5 24 Apr 02 12 10 0 0 2
18 3 107 2 26 Apr 02 15 9 4 2 0
19 16 27 3 26 Apr 02 10 9 1 0 0
20 1 159 1 27 Apr 02 6 6 0 0 0
21 13 41 2 28 Apr 02 6 6 0 0 0
22 7 131 4 29 Apr 02 8 6 0 2 0
23 9 83 6 30 Apr 02 7 6 0 0 1
24 2 79 3 30 Apr 02 9 9 0 0 0
25 8 128 5 30 Apr 02 12 10 1 1 0
26 11 169 4 30 Apr 02 11 10 0 0 1
Total 253 205 16 24 8

a Sow sick at farrowing.

Item Value Target

Percent of litters with scours in 12 months before outbreak 7% < 5%
Preweaning mortality in 12 months before outbreak 11.5% < 12%
Post weaning mortality 5% < 3%
Gross margin per pig marketed $35.00 -
Treatment costs per litter $10.00 -
E. coli vaccine 2 × $2.50 -
Labour cost to vaccinate one pig $0.30 -

102
12 Review questions

12.1 Host, agent, environment

You are discharging a 2 year-old male domestic shorthair cat who has spent 10 days in your clinic
recovering from the complications associated with obstruction of the urinary tract. As the cat’s owner
is writing out a cheque for $1500 he asks ‘will my cat experience another attack of FUS in the future
and what can I do to prevent it?’ What advise would you give, from an epidemiological perspective?
Think about three or four health problems or diseases that you or your friends have had. List each of
the host, agent, and environmental factors that may have been causative for each disease you have
listed.
Can you think if circumstances when exposure to a causal factor does not change disease inci-
dence?
List five or six broad and fundamental influences on health and disease, that is, those influences
that change the population patterns of disease.
Reflect on some medical and public health activities which were widely practiced but are now known
to be wrong, some dangerously so. Your reflection should include some historical activities say,
before the turn of the twentieth century and more recent ones. Also reflect on some current policies
and practices that may meet the same fate.

12.2 Measures of health

Imagine you are in a country where no animal demographic data is available. An epidemic of pneu-
monia is suspected in the cattle population. You are asked to develop a plan to prevent and control
the epidemic. Which questions do you need to answer to start a rational control strategy for this
disease? Which epidemiological data do you need to answer the questions?
What benefits are there from investigating the changes in disease frequency in a population over
time?
Consider the reasons why a variation in disease pattern might be artefact rather than real. Can you
group them into three or four categories of explanation? What explanations can you think of for a
real change in disease frequency? Can you group these into three or four categories of explanation?
Imagine you are asked to describe the health status of a population of animals to a senior public
servant. The person you are talking to has no previous background in animal (or human) health.
What kinds of measures would you choose to portray the health of the animal population? Consider
not only the specific types of data, but also the qualities of the data you would seek out.
Imagine a population of 10,000 new army recruits. You are interested in studying the incidence and
prevalence rate of gunshot wounds on war duty. Assume all gunshot wounds lead to permanent
visible damage. You follow the recruits for one year. All of the study population survive, all medical
records are available, and all recruits are available to interview and examination. Assume the oc-
currence of gunshot wounds is spread evenly through the year, and that at the time of entering the
army, no recruits had gunshot wounds. Over the year you determine that 20 recruits had a gunshot
wound.

103
 • What is the incidence risk of gunshot wounds? What is the incidence rate of gunshot wounds?

• What is the point prevalence rate of having had a gunshot wound at the beginning, middle,
and end of the year?

• What is the period prevalence rate over the year?

• If the incidence rate remains the same over time, what is the prevalence rate of ever being
scarred by the end of five years?

• What is the average duration of a gunshot wound, among those scarred, by the end of the first
year?

• What is the estimated point prevalence rate over the five-year period?

What might be your denominator for a study defining the incidence rate of:

• Calf mortality.

• Clinical mastitis.

• Bovine spongiform encephalopathy.

12.3 Measures of association

Reflect on the terms ‘risk factor’ and ‘cause of disease.’ What is the difference between these terms?
Consider why the risk ratio might provide a false picture of the effect of a risk factor on disease and
hence the strength of association.
Imagine that the incidence of chronic obstructive pulmonary disease (COPD) in horse is compared
in two areas of a country: one with polluted air (A) and the other not (B). In the polluted area there
were 20 cases of COPD in a population of 100,000. In the other area there were 10 cases in a
population of 100,000.

• What is the risk ratio of COPD in area (A)?

• What is the risk ratio of COPD in area (B)?

• Do we know the precision of these estimates of risk ratio?

• What explanations are there for the risk ratio estimate in area (A)?

• What questions will you need to consider before concluding that there is a real association
between pollution and COPD?

104
Imagine that exposure to a dry cat food triples the incidence of a feline urologic syndrome (FUS),
that is, the risk ratio is 3. This disease has a baseline incidence of 1 per cent per year in the non-
exposed group. Imagine also that the baseline incidence is double in castrated male cats (that is,
2 per cent) and that the risk ratio associated with exposure to dry cat food is the same, three. You
follow 100 entire and 100 castrated male cats that are fed dry cat food, and an equivalent number of
cats fed moist food. The study lasts for 5 years. Create a 2 × 2 table to show the data for castrates
and entire male cats and calculate the odds ratio of disease in the exposed group in relation to those
not exposed. Compare the odds ratio with the risk ratio of 3.
The Ministry of Health has made available a sum of $100,000 for a health promotion programme to
reduce coronary heart disease mortality. We can spend it on encouraging people to stop smoking or
encouraging them to do more exercise. Assume the risk ratio associated with both risk factors is 2,
that changes in prevalence rate are equally permanent, and that the cardioprotective effect occurs
quickly. Which choice will give a better return in lives saved?

• First, make a judgement on which of the two preventive programs you prefer.
• Now consider which is more common: smoking or lack of exercise?
• Calculate the population attributable risk when the prevalence rate of smoking is 20%, 30%,
40% and 50% and the prevalence rate of lack of exercise is 60%, 70%, and 80% (these are
realistic prevalence rates in industrialised countries). Has the result altered or substantiated
your earlier judgement?

12.4 Study design

Imagine a cohort study which aims to determine the incidence of arthritis in large breeds of dogs.
The follow-up period for the study is five years. Describe the advantages and disadvantages of the
two approaches for measuring incidence.
Imagine a study of the incidence of congestive heart disease in large breeds of dogs, base on post
mortem records collected at a University teaching hospital over a five-year period. Again, consider
the advantages and disadvantages of the two approaches for measuring incidence.
Is there a difference between a clinical case series and a population case series?
How might epidemiology study the potential role in disease causation of factors which vary little
between individuals within a region or country. For example: fluoride content of water, hardness or
softness of water supplies, annual exposure to sunshine?
What is the essential feature that differentiates a cross-sectional study from a cohort study?
Explain what you understand by the term ‘error’. What is the difference, if any, between error and
bias?

12.5 Diagnostic tests

A client of your manages a study beef herd which, for the past ten years, has consistently tested
negative for tuberculosis. A positive reactor has been found after the latest round of testing. What
would you advise?

105
13 Resources

EpiCentre, Massey University http://epicentre.massey.ac.nz/

University of Guelph, Department of Pop Medicine http://www.ovc.uoguelph.ca/popm/
Royal Veterinary College, University of London http://www.rvc.ac.uk/
University of Michigan School of Public Health http://www.sph.umich.edu/epid/

Epidemiology Monitor http://www.epimonitor.net/

Association of Teachers of Veterinary Public Health http://www.cvm.uiuc.edu/atvphpm/
Centers for Disease Control and Prevention http://www.cdc.gov/
EXCITE http://www.cdc.gov/excite/
Epidemiology Supercourse http://www.pitt.edu/~super1/
VEIN links: Evidence Based Medicine http://vein.library.usyd.edu.au
Centre for Veterinary Education, University of Sydney http://www.cve.edu.au/
EBM Resources http://www.dartmouth.edu/~biomed/resources
.htmld/guides/ebm_resources.shtml

MAF, New Zealand http://www.maf.govt.nz

DAFF, Australia http://www.daff.gov.au/
Canadian Food Inspection Agency http://www.inspection.gc.ca
Health Canada http://www.hc-sc.gc.ca/
Instituto Nacional de Tecnología Agropecuaria http://www.inta.gov.ar/
The Cochrane Collaboration http://www.cochrane.org

106
14 Epdemiology formulae sheet

14.1 Measures of association

Disease event data presented in a 2 × 2 table format:

Diseased Non-diseased Total

Exposed a b a+b
Non-exposed c d c+d
Total a+c b+d a+b+c+d=n

Cross-sectional and cohort studies:

Incidence risk of disease in the exposed: RE+ = a/(a + b)
Incidence risk of disease in the non-exposed: RE− = c/(c + d)
Incidence risk in the total population: Rtotal = a + c/n
Odds of disease in the exposed: OE+ = a/b
Odds of disease in the non-exposed: OE− = c/d

Case-control studies:
Odds of exposure in cases: OD+ = a/c
Odds of exposure in controls: OD− = b/d

107
Odds ratio (OR) cohort studies: Odds ratio (OR) case-control studies:

OE+ OD+
OR = OR =
OE− OD−

Incidence risk ratio (RR):

RE+
RR =
RE−

Attributable risk (AR): Attributable fraction (AF):

AR = RE+ − RE− (RE+ − RE− )

AF =
RE+

Population attributable risk (PAR): Population attributable fraction (PAF):

PAR = Rtotal − RE− (Rtotal − RE− )

PAF =
Rtotal

108
14.2 Diagnostic tests

Diagnostic test data presented in a 2 × 2 table format:

Disease positive Disease negative Total

Test positive a b a+b
Test negative c d c+d
Total a+c b+d a+b+c+d=n

True prevalence (TP): Apparent prevalence (AP):

(a + c) (a + b)
TP = AP =
n n

Sensitivity (Se): Specficity (Sp):

a d
Se = Sp =
(a + c) (b + d)

Positive predictive value (PPV): Negative predictive value (NPV):

a d
PPV = NPV =
(a + b) (c + d)

Estimating true prevalence from apparent prevalence (after Rogan and Gladen 1978):

AP + Sp − 1
TP =
Se + Sp − 1

109
Likelihood ratios — Bayes nomogram

Positive likelihood ratio (LR+ ): Negative likelihood ratio (LR− ):

Se 1 − Se
LR+ = LR− =
1 − Sp Sp

110
14.3 Sampling

Sampling to estimate population parameters

Formulae used to determine the required sample size to estimate population parameters (population
total, mean, and proportion) on the basis of a simple random sample.

z 2 SD2 z 2 SD2 z 2 (1 − Py )Py

Total: n > Mean: n > Proportion: n >
2 2 2

Where:
z : the reliability coefficient (e.g. z = 1.96 for an alpha level of 0.05).
SD: the population standard deviation of the variable of interest.
Py : the unknown population proportion.
: the maximum absolute difference between the sample estimate and the unknown population
value.
Sampling to detect the presence of disease
Formula to calculate an appropriate sample size to detect the presence of disease:

1 D−1
n = (1 − α D ) × (N − )
2
Where:
N : the population size.
α: 1 - confidence level (usually α = 0.05).
D: the estimated minimum number of diseased animals in the group (that is, population size × the
minimum expected prevalence).

111
References
Altman, D., & Bland, J. (1994a). Statistics Notes: Diagnostic tests 1: sensitivity and specificity.
British Medical Journal, 308, 1552.

Altman, D., & Bland, J. (1994b). Statistics notes: Diagnostic tests 2: predictive values. British
Medical Journal, 309, 102.

Ast, D., & Schlesinger, E. (1956). The conclusion of a ten-year study of water fluoridation. American
Journal of Public Health, 46, 265 - 271.

Brenner, H., Greenland, S., & Savitz, D. (1992). The effects of nondifferential confounder misclassi-
fication in ecological studies. Epidemiology , 3, 456 - 469.

Carey, J., Klebanoff, M., Hauth, J., Hillier, S., Thom, E., & Ernest, J. (2000). Metronidazole to prevent
preterm delivery in pregnant women with asymptomatic bacterial vaginosis. New England Journal
of Medicine, 342, 534 - 540.

Dawson, B., & Trapp, R. (2004). Basic and Clinical Biostatistics. New York: McGraw-Hill Medical.

Deeks, J., & Altman, D. (2004). Statistics Notes: Diagnostic tests 4: likelihood ratios. British Medical
Journal, 329, 168 - 169.

Dohoo, I., Martin, S., & Stryhn, H. (2003). Veterinary Epidemiologic Research. Charlottetown,
Prince Edward Island, Canada: AVC Inc.

Donnelly, C., Ghani, A., Leung, G., Hedley, A., Fraser, C., Riley, S., et al. (2004). Epidemiological
determinants of spread of causal agent of severe acute respiratory syndrome in Hong Kong.
Lancet, 361, 1761 - 1766.

Draper, G., Vincent, T., Kroll, M., & Swanson, J. (2005). Childhood cancer in relation to distance
from high voltage power lines in England and Wales: a case-control study. British Medical Journal,
330, 1290.

Elwood, J. (2007). Critical Appraisal of Epidemiological Studies and Clinical Trials. New York, USA:
Oxford University Press.

Farquharson, B. (1990). On-farm trials. In Epidemiological Skills in Animal Health. Refresher Course
for Veterinarians. Proceedings 143 (p. 207 - 212). Postgraduate Committee in Veterinary Science,
University of Sydney, Sydney, Australia.

Fletcher, R., Fletcher, S., & Wagner, E. (1996). Clinical Epidemiology. Baltimore, USA: Williams
and Wilkins.

Fosgate, G., & Cohen, N. (2008). Review Article. Epidemiological study design and the advance-
ment of equine health. Equine Veterinary Journal, 40(7), 693 - 700.

Fransen, M., Woodward, M., Norton, R., Robinson, E., Butler, J., & Campbell, A. (2002). Excess
mortality or institutionalisation following hip fracture: men are at greater risk than women. Journal
of the American Geriatrics Society , 50, 685 - 690.

112
Fraser, D., Tsai, T., Orenstein, W., Parkin, W., Beecham, H., Sharrar, R., et al. (1977). Legionnaires’
disease — description of an epidemic of pneumonia. New England Journal of Medicine, 296,
1189-1197.

Friss, R., & Sellers, T. (2009). Epidemiology for Public Health Practice. New York, USA: Jones and
Bartlett.

Gardner, I. (1990a). Case study: Investigating neo-natal diarrhoea. In D. Kennedy (Ed.), Epidemi-
ology at Work. Refresher Course for Veterinarians. Proceedings 144 (p. 109 - 129). Quarantine
Station, North Head, NSW: Postgraduate Committee in Veterinary Science, University of Sydney,
Sydney, Australia.

Gardner, I. (1990b). Reporting disease outbreaks. In D. Kennedy (Ed.), Epidemiology at Work.

Refresher Course for Veterinarians. Proceedings 144 (p. 29 - 42). Quarantine Station, North
Head, NSW: Post Graduate Committee in Veterinary Science, The University of Sydney, Sydney,
Australia.

Gardner, L., Landsittel, D., & Nelson, N. (1999). Risk factors for back injury in 31,076 retail mer-
chandise store workers. American Journal of Epidemiology , 150, 825 - 833.

Gerstman, B. (2003). Epidemiology Kept Simple: An Introduction to Traditional and Modern Epi-
demiology. New York, USA: John Wiley and Sons.

Goodwin-Ray, K., Stevenson, M., & Heuer, C. (2008c). Flock-level case–control study of slaughter-
lamb pneumonia in New Zealand. Preventive Veterinary Medicine, 85, 136 - 149.

Greenhalgh, T. (1997a). How to read a paper: Assessing the methodological quality of published
papers. British Medical Journal, 315, 305 - 308.

Greenhalgh, T. (1997b). How to read a paper: Getting your bearings (deciding what the paper is
about). British Medical Journal, 315, 243 - 246.

Greenhalgh, T. (1997c). How to read a paper: Papers that report diagnostic or screening tests.
British Medical Journal, 315, 540 - 543.

Greenhalgh, T. (1997d). How to read a paper: Papers that summarise other papers (systematic
reviews and meta-analyses). British Medical Journal, 315, 672 - 675.

Greenhalgh, T. (1997e). How to read a paper: Papers that tell you what things cost (economic
analyses). British Medical Journal, 315, 596 - 599.

Greenhalgh, T. (1997f). How to read a paper: Statistics for the nonstatistician. II: ‘Significant’
relations and their pitfalls. British Medical Journal, 315, 422 - 425.

Greenhalgh, T. (1997g). How to read a paper: The Medline database. British Medical Journal, 315,
180 - 183.

Greenhalgh, T. (2006). How to Read a Paper: The Basics of Evidence-Based Medicine. London:
British Medical Journal Books.

Greenhalgh, T., & Taylor, R. (1997). How to read a paper: Papers that go beyond numbers (qualita-
tive research). British Medical Journal, 315, 740 - 743.

113
Greenland, S. (2009). Interactions in epidemiology: Relevance, identification, and estimation. Epi-
demiology , 20, 14 - 16.

Hill, A. (1965). The environment and disease: Association or causation? Proceedings of the Royal
Society of London. Series C, Medicine, 58, 295 - 300.

Hoyert, D., Arias, E., Smith, B., Murphy, S., & Kochanek, K. (1999). Deaths: final data for 1999.
National Vital Statistics Reports Volume 49, Number 8. Hyattsville MD: National Center for Health
Statistics.

Johansen, C., Boise, J., McLaughlin, J., & Olsen, J. (2001). Cellular telephones and cancer — a
nationwide cohort study in Denmark. Journal of the National Cancer Institute, 93, 203 - 237.

Kelsey, J., Thompson, W., & Evans, A. (1996). Methods in Observational Epidemiology. London:
Oxford University Press.

Knol, M., Tweel, I. van der, Grobbee, D., Numans, M., & Geerlings, M. (2007). Estimating interaction
on an additive scale between continuous determinants in a logistic regression model. International
Journal of Epidemiology , 36(5), 1111 - 1118.

Leung, W.-C. (2002). Measuring chances. Student British Medical Journal, 10, 268 - 270.

Levy, P., & Lemeshow, S. (1999). Sampling of Populations Methods and Applications. London:
Wiley Series in Probability and Statistics.

Mackintosh, C., Schollum, L., Harris, R., Blackmore, D., Willis, A., Cook, N., et al. (1980). Epidemi-
ology of leptospirosis in dairy farm workers in the Manawatu. Part I: A cross-sectional serological
survey and associated occupational factors. New Zealand Veterinary Journal, 28, 245 - 250.

Martin, S., Meek, A., & Willeberg, P. (1987). Veterinary Epidemiology Principles and Methods.
Ames, Iowa: Iowa State University Press.

Merrill, R., & Timmreck, T. (2006). Introduction to Epidemiology (4th ed.). San Bernardino: California
State University.

Morris, R. (1990). Disease outbreak! What can you do? In Epidemiological Skills in Animal Health.
Refresher Course for Veterinarians. Proceedings 143 (p. 321 - 327). Postgraduate Committee in
Veterinary Science, University of Sydney, Sydney, Australia.

Muscat, J., Malkin, M., Thompson, S., Shore, R., Stellman, S., & McRee, D. (2000). Handheld
cellular telephone use and risk of brain cancer. Journal of the American Medical Association,
284, 3001 - 3007.

Noordhuizen, J., Frankena, K., Hoofd, C. van der, & Graat, E. (1997). Application of Quantitative
Methods in Veterinary Epidemiology. Wageningen: Wageningen Pers.

Oleckno, W. (2002). Essential Epidemiology Principles and Applications. Prospect Heights, Illinois:
Waveland Press.

Olsen, J. (2003). What characterises a useful concept of causation in epidemiology? Journal of

Epidemiology and Community Health, 57 , 86 - 88.

114
Parsonnet, J., Friedman, G., Vandersteen, D., Chang, Y., Vogelman, J., Orentreich, N., et al. (1991).
Helicobacter pylori infection and the risk of gastric-carcinoma. New England Journal of Medicine,
325(16), 1127 - 1131.

Petrie, A., & Watson, P. (2005). Statistics for Veterinary and Animal Science. London: Blackwell
Science.

Pfeiffer, D., Robinson, T., Stevenson, M., Stevens, K., Rogers, D., & Clements, A. (2008). Spatial
Analysis in Epidemiology. New York, USA: Oxford University Press.

Porta, M., Greenland, S., & Last, J. (2008). A Dictionary of Epidemiology. New York, USA: Oxford
University Press.

Rinzin, K., Stevenson, M., Probert, D., Bird, R., Jackson, R., French, N., et al. (2008). Free-roaming
and surrendered dogs and cats submitted to a humane shelter in Wellington, New Zealand, 1999
– 2006. New Zealand Veterinary Journal, 56, 297 - 303.

Rogan, W., & Gladen, B. (1978). Estimating prevalence from results of a screening test. American
Journal of Epidemiology , 107 , 71 - 76.

Rothman, K. (1976). Causes. American Journal of Epidemiology , 104, 587 - 592.

Rothman, K., Greenland, S., & Lash, T. (2008). Modern Epidemiology. Philadelphia, USA: Lippin-
cott, Williams and Wilkins.

Schlesselman, J. (1982). Case-Control Studies Design, Conduct, Analysis. London: Oxford Univer-
sity Press.

Schwarz, D., Grisso, J., Miles, C., Holmes, J., Wishner, A., & Sutton, R. (1994). A longitudinal
study of injury morbidity in an African-American population. Journal of the American Medical
Association, 271, 755 - 760.

Scuffham, A., Legg, S., Firth, E., & Stevenson, M. (2009). Prevalence and risk factors for muscu-
loskeletal discomfort in New Zealand veterinarians. Applied Ergonomics, 41, 444 - 453.

Selvin, S. (1996). Statistical Analysis of Epidemiological Data. London: Oxford University Press.

Siscovick, D., Weiss, N., Fletcher, R., & Lasky, T. (1984). The incidence of primary cardiac-arrest
during vigorous exercise. New England Journal Of Medicine, 311, 874 - 877.

Smith, R. (1995). Veterinary Clinical Epidemiology — A Problem-Oriented Approach. Boca Raton,

Florida: CRC Press.

Stevenson, M., Morris, R., Lawson, A., Wilesmith, J., Ryan, J., & Jackson, R. (2005). Area-level risks
for BSE in British cattle before and after the July 1988 meat and bone meal feed ban. Preventive
Veterinary Medicine, 69, 129 - 144.

Stevenson, M., Wilesmith, J., Ryan, J., Morris, R., Lawson, A., Pfeiffer, D., et al. (2000). Descriptive
spatial analysis of the epidemic of bovine spongiform encephalopathy in Great Britain to June
1997. Veterinary Record, 147 , 379 - 384.

Thrusfield, M. (2007). Veterinary Epidemiology. London: Blackwell Science.

115
Trivier, J., Caron, J., Mahieu, M., Cambier, N., & Rose, C. (2001). Fatal aplastic anaemia associated
with clopidogrel. Lancet, 357 , 446.

Valent, F., Brusaferro, S., & Barbone, F. (2001). A case-crossover study of sleep and childhood
injury. Pediatrics, 107 , E23.

Vander Stoep, A., Beresford, S., & Weis, N. (1999). A didactic device for teaching epidemiology
students how to anticipate the effect of a third factor on an exposure-outcome relation. American
Journal of Epidemiology , 150, 221.

Webb, P., Bain, C., & Pirozzo, S. (2005). Essential Epidemiology. Cambridge, UK: Cambridge
University Press.

Wilesmith, J., Stevenson, M., King, C., & Morris, R. (2003). Spatio-temporal epidemiology of foot-
and-mouth disease in two counties of Great Britain in 2001. Preventive Veterinary Medicine, 61,
157 - 170.

Will, R., Ironside, J., Zeidler, M., Cousens, S., Estibeiro, K., & Alperovitch, A. (1996). A new variant
of Creutzfeld-Jacob disease in the UK. Lancet, 347 , 921 - 925.

Yang, C., Chiu, H., Cheng, M., & Tsai, S. (1998). Chlorination of drinking water and cancer in
Taiwan. Environmental Research, 78, 1 - 6.

116