Week 1
How do we approach
epidemiology investigation?
In general, how do we
approach epidemiology investigation?
We perform ‘studies’. we are interested in causation!
What is a Study?
- A STUDY IS A MEASUREMENT DEVICE ‐ analogous to a scale of weights or
a measuring tape
- If the study involves comparing the occurrence of events among human
populations, it is an epidemiologic study
- Epidemiology,
however, does not determine the cause of a disease in a given
individual!
If a study is a
measurement device, what are the measurement units?
- Since we are concerned with relating exposure to the occurrence of
disease, we assess outcome in a population by using a measure of
occurrence; basic measures of
occurrence are incidence rates, cumulative incidence, and
prevalence
- The object of measurement may be a rate or a risk, but typically it
is a measure of effect (or
association) to describe the difference in outcome occurrence
among exposure groups; attributable risk (AR or RD, the difference
between measures of occurrence) or relative risk (RR, the ratio of the
measures of occurrence) are basic measures of association
- The final result of a study can be expressed as a single number (in
measure‐of‐ association units) or
‘POINT ESTIMATE’; this is the
best estimate the study provides as to the size of the thing being
measured,
What is the goal of
the study?
- In common with other measuring devices, the study has a simple goal
‐ AN ACCURATE MEASUREMENT
- ACCURACY, which means the degree to which the measurement reflects
the true state of the universe, is comprised of VALIDITY and PRECISION
- Hence, a study can be valid or precise, both, or neither.
What do we mean by
validity?
- A VALID STUDY HAS LITTLE
‘SYSTEMATIC ERROR’ OR ‘BIAS’
- Bias refers to the tendency of a measurement to deviate from the
truth in the same direction, systematically
two types of
validity?
- Internal validity ‐
is the notion of systematic error, refers to the measurements made
within the study population
- External validity ‐
refers to the applicability of the measurements made from the study
population to larger, potentially more diverse target populations
- Internal validity is a
prerequisite for external validity
- If the measurement you’ve made on the actual people in the study is
too biased, there’s not much point in worrying about unstudied groups to
whom the measurement might generalize.
What do we mean by
precision?
- A PRECISE STUDY HAS LITTLE ‘RANDOM ERROR’ AND PRECISION, RELIABILITY
AND REPRODUCIBILITY REFER TO THE SAME THING
- Chance refers to
measurements that deviate from the truth in any direction,
randomly
How to
conceptualize random vs. systematic error?

- Reliable but not valid
- Valid but not reliable
- Both reliable and valid
‘In a sufficiently large study, virtually all errors of
concern are systematic errors’ ‐ Rothman 
What are the
possible explanations for the measurement result a study gives?
THE MEASUREMENT IS EXPLAINED BY ONLY THREE FACTORS:
Etiology – why some
study designs are better than others!
This usually requires that we go beyond group association and
establish three definitive requirements (lung cancer example):
The “cause” is associated with the “effect” at the individual
level The potential “cause” and the potential “effect” occur more
frequently at the INDIVIDUAL level than would be expected by chance
E.g., we establish that individuals with lung cancer are more
frequently smokers than individuals without lung cancer
The “cause” precedes the “effect” in time (temporality) The
potential “cause” is PRESENT AT AN EARLIER TIME than the potential
“effect” E.g., we establish that cigarette smoking comes before the
development of lung cancer Altering the “cause” alters the “effect”
When the potential “cause” is reduced or eliminated, the potential
“effect” is also reduced or eliminated – OFTEN INVOLVES EXPERIMENTAL
DESIGN E.g., we establish that reducing cigarette smoking reduces lung
cancer rate
Bias/error and/or confounding
Chance
Cause (provided that the study results are valid and precise)
- The simplest meaning of cause is that we infer (derive, surmise or
suggest) that an exposure ‘E’ precedes a disease outcome ‘D’
E ———————> D
- In designing a study we strive to improve validity by minimizing the
potential role of bias and confounding, and to improve precision by
minimizing the potential role of chance
In modern causal inference, we often try to gauge how well the study
has succeeded in doing these two things
IMPORTANTLY, SOME STUDY
DESIGNS ARE BETTER THAN OTHERS (RCTs > Cohort > Case‐Control >
Ecological) – WHY?
Etiology – why some
study designs are better than others!
This usually requires that we go beyond group association and
establish three definitive requirements (lung cancer example):
1. The “cause” is associated
with the “effect” at the individual level - The potential “cause”
and the potential “effect” occur more frequently at the INDIVIDUAL level
than would be expected by chance - E.g., we establish that individuals
with lung cancer are more frequently smokers than individuals without
lung cancer
- The “cause” precedes the
“effect” in time (temporality)
- The potential “cause” is PRESENT AT AN EARLIER TIME than the
potential “effect”
- E.g., we establish that cigarette smoking comes before the
development of lung cancer
- Altering the “cause” alters
the “effect”
- When the potential “cause” is reduced or eliminated, the potential
“effect” is also reduced or eliminated – OFTEN INVOLVES EXPERIMENTAL
DESIGN
- E.g., we establish that reducing cigarette smoking reduces lung
cancer rate
How we use studies
to establish etiology?

What are the major
types of bias?
- Selection bias
Usually occurs when the
association between exposure and disease differs for those who
participate in study vs. non‐participants:
Examples:
- Volunteer bias
- Healthy worker effect
- Lost to follow‐up bias
- Overmatching (controls not selected independent of exposure)
- Surveillance or detection bias
- Collider stratification bias
Can be DIFFERENTIAL or NON‐DIFFERENTIAL:
Differential
: usually exposure differs according to outcome, e.g., recall
bias – can bias in either direction and more unpredictable
Non‐differential:
unrelated to other study variables, e.g., recall limitation ‐ tends to
produces estimates closer to null value and more predictable
- Potential solutions:
Make accurate measurements of exposure and outcome variables
‐ an example is the use of blinding methods
- Randomization
- Select subjects carefully and keep them on the study
Selection bias: One relevant group in the population (exposed
cases in the example) has a higher probability of being included in the
study sample.