EPB PHC 6000 EPIDEMIOLOGY FALL, 1997

1
Unit 5 Evaluating Associations
2

• Unit 5 Learning Objectives
• Understand the concepts of valid and causal
  statistical associations.
• Understand the possible explanations for
  statistical associations
• --- Chance
• --- Bias
• --- Confounding
• Distinguish between the major types of bias in
  epidemiologic studies.
• Distinguish between internal validity and
  external validity.

3

• Unit 5 Learning Objectives (cont.)
• Understand the concept of confounding and methods
  to assess its presence.
• Understand the concept of causality, including
  necessary and sufficient causes.
• Understand pros and cons of guidelines used to
  evaluate causal associations in epidemiology.

4
Assigned Readings Textbook (Gordis)
Chapter 14, pages 206-222 (from association to
causation) Chapter 15, pages 224-230 (bias and
confounding)
5
Evaluating Associations
If we observe an exposure/disease association, we
must consider 1. Is the association valid? (do
the study findings reflect the true
relationship between the exposure and
disease?) 2. Is the association causal? (Is
there sufficient evidence to infer
that a causal association exists
between the exposure and the disease?)
6
Evaluating Associations
EVALUATING THE VALIDITY OF AN ASSOCATION In any
epidemiologic study, there are at least
3 possible explanations for the observed
results 1. CHANCE 2. BIAS 3.
CONFOUNDING These explanations are not mutually
exclusive -- more than one can be present in the
same study
7
CHANCE
The luck of the draw
1. Rarely can we study an entire population,
so inference is attempted from a sample of the
population 2. There will always be random
variation from sample to sample 3. In general,
smaller samples have less precision,
reliability, and statistical power (more
sampling variability)
8
CHANCE
4. A test of statistical significance is
performed to assess the degree to which
the data are compatible with the null
hypothesis of no association 5. The
p-value reflects the probability that
the test statistic (e.g. t-statistic or chi-
square statistic) produced from the data
would be greater than or equal to the
observed value under the null hypothesis
of no association
9
CHANCE
6. By convention, if p lt 0.05, then the
association between the exposure and
disease is considered to be statistically
significant. (e.g. we reject the null
hypothesis (H0) and accept the alternative
hypothesis (H1)) What does p lt 0.05
mean? Indirectly, it means that we suspect that
the magnitude of effect observed (e.g. risk
ratio) is not due to chance alone (in the
absence of biased data collection or
analysis)
10
CHANCE
Example Possible biased coin Coin Toss - 10
Times Observed Expected Heads 7 5 Ta
ils 3 5 Odds HT 73 2.33 Excess Heads O
- E 7 - 5 2 p-value gt 0.05 (computation not
shown) The observed excess of heads to tails is
not much greater than that which might be
expected by chance
11
CHANCE
Example Possible biased coin Coin Toss 1,000
Times Observed Expected Heads 700 500
Tails 300 500 Odds HT 700300
2.33 Excess Heads O - E 700 - 500
200 p-value lt 0.05 (computation not shown) The
observed excess of heads to tails is much greater
than that which might be expected by chance
12
CHANCE
BEWARE The p-value reflects both the magnitude
of the difference between the study groups AND
the sample size So, a related, but more
informative measure known as the confidence
interval (CI) is also calculated. CI a range
of values within which the true population value
falls, with a certain degree of assurance
(probability).
13
CHANCE
HYPOTHETICAL EXAMPLE OF 95 CONFIDENCE
INTERVAL Exposure Caffeine intake (high versus
low) Outcome Incidence of breast cancer Risk
Ratio 1.32 (point estimate) p-value 0.14
(not statistically significant) 95 C.I. 0.87
- 1.98 __________________________________________
___ 0.0 0.5 1.0 1.5 2.0 (null value)
95 confidence interval
14
CHANCE
INTERPRETATION Our best estimate is that women
with high caffeine intake are 1.32 times (or 32)
more likely to develop breast cancer compared to
women with low caffeine intake. However, we are
95 confident that the true value (risk) of the
population lies between 0.87 and 1.98 (assuming
an unbiased study). ____________________________
_________________ 0.0 0.5 1.0 1.5 2.0
(null value)
95 confidence interval
15
CHANCE
So, if the 95 confidence interval does NOT
include the null value of 1.0 (p lt 0.05), then we
declare a statistically significant
association. Question How do we declare a
valid statistical association? Answer By
ruling out, to the extent possible, bias or
confounding
16
CHANCE
Note Although we have initially reviewed the
role of chance, the conventional sequence would
be to first assess (rule out) the presence of
systematic bias. In other words, it makes no
sense to evaluate chance as a possible
explanation when a study is biased from the start
(i.e. the experiment we have set up is flawed,
and hence, we should discard it).
17
BIAS

• BIAS Systematic error in the design, conduct,
  or analysis of a study that results in a mistaken
  estimate of an exposure/disease relationship
• SELECTION BIAS
• INFORMATION BIAS
• Interviewer
• Recall Bias
• Reporting Bias
• Surveillance Bias

18
BIAS
SELECTION BIAS Any systematic error that arises
in the process (mechanism) of identifying the
study populations (i.e. the two 2 study groups to
be compared) Occurs when selection of study
subjects (whether by exposure or disease status)
is based on different criteria related to
exposure or disease Results in the study
groups being non-comparable, unless some type of
statistical adjustment can be made
19
SELECTION BIAS
EXAMPLE Case Control Study Outcome Hemorrhagi
c stroke Exposure Appetite suppressant products
that contain Phenylpropanolamine
(PPA) Cases Persons who experienced a
stroke Controls Persons in the community without
stroke Bias Control
subjects were recruited by random-digit
dialing from 900 AM to 500 PM. This
resulted in over- representation of unemployed
persons who may not represent the study base
in terms of use of appetite
suppressant products.
20
SELECTION BIAS
EXAMPLE Retrospective Cohort
Study Outcome COPD Exposure Employment in
tire
manufacturing Exposed Plant assembly line
workers Non-exposed Plant administrative
personnel Bias The exposed were contacted
(selected) at a local pub while watching Monday
night football the non-exposed were identified
through review of plant personnel files. Exposed
persons may have been more likely to be smokers
(related to COPD)
21
SELECTION BIAS
EXAMPLE Non-Response If refusal or
non-response is related to exposure, the estimate
of effect (exposure/disease) may be biased. For
example, if controls are selected by use of a
household survey, non-response may be related to
demographic and lifestyle factors associated with
employment. Responders often differ
systematically from persons who do not respond.
22
SELECTION BIAS
NOTE Restrictive sampling alone, so long as
different criteria are not used between study
groups, does not confer selection bias It
merely means that the study results may not
generalize to the larger population (external
validity)
23
INFORMATION BIAS
Definition Systematic differences in the way in
which data on exposure and outcome are obtained
from the various study groups. Some
Types/Sources of Information Bias Bias in
abstracting records Bias in interviewing Bias
from surrogate interviews Surveillance
bias Reporting and recall bias
24
INTERVIEWER BIAS
DEFINITION Systematic difference in the
soliciting, recording, or interpretation of
information from study participants Can affect
every type of epidemiologic study May occur
when interviewers are not blinded to exposure
or outcome status of participants.
25
INTERVIEWER BIAS
Interviewers knowledge of subjects disease
status may result in differential probing of
exposure history Similarly, interviewers
knowledge of subjects exposure history may
result in differential probing and recording of
the outcome under examination Placebo control
is one method used to maintain observer blindness
in randomized trials.
26
RECALL BIAS
DEFINITION Study group participants
systematically differ in the way data on exposure
or outcome are recalled Particularly
problematic in case-control studies Individuals
who have experienced a disease or adverse health
outcome may tend to think about possible causes
of the outcome. This can lead to differential
recall
27
RECALL BIAS
EXAMPLE Case-Control Study Outcome Cleft
palate Exposure Systemic infection during
pregnancy Cases Mothers giving birth to
children with cleft palate Controls Mothers
giving birth to children free
of cleft palate Bias Mothers who have given
birth to a child with cleft palate may recall
more thoroughly colds and other infections
experienced during pregnancy
28
REPORTING BIAS
DEFINITION Selective suppression or revealing
of information such as past history of sexually
transmitted disease. Often occurs because
subject reluctance to report an exposure due to
attitudes, beliefs, and perceptions Wish
bias may occur among subjects who have developed
a disease and seek to show that the disease is
not their fault.
29
SURVEILLANCE BIAS
If a population is monitored over a period of
time, disease ascertainment may be better in the
monitored population than in the general
population (surveillance bias). May lead to
biased estimate of exposure/disease relationship.
30
MISCLASSIFICATION
DEFINITION Erroneous classification of the
exposure or disease status of an individual into
a category to which it should not be
assigned Example --- Cases incorrectly
classified as controls --- Controls incorrectly
classified as cases --- Exposed incorrectly
classified as non- exposed --- Non-expo
sed incorrectly classified as exposed
31
MISCLASSIFICATION
Non-differential misclassification The
proportion of subjects misclassified on exposure
does not depend on disease status OR The
proportion of subjects misclassified on disease
does not depend on exposure status
32
MISCLASSIFICATION
Non-differential misclassification Tends to
make the exposure or disease groups more similar
than they really are Some non-differential
misclassification is inevitable Almost always
results in bias towards the null In
interpretation, researcher must consider what
real effect might have been obscured
33
MISCLASSIFICATION
Differential misclassification Classification
error of exposure status occurs more frequently
among the diseased or non-diseased OR Classifica
tion error of disease status occurs more
frequently among the exposed or non-exposed
34
MISCLASSIFICATION
Differential misclassification Results in
relatively unpredictable effects Can
exaggerate or underestimate the true
exposure/disease relationship By chance
(infrequently), can also result is estimate that
is the same as the true exposure/disease
relationship.