Ramona Dumitrescu

r.dumitrescu@saba.edu
Saba University Medical School
Spring 2013
Descriptive data provide the first clues in the
investigation of a cause-effect relationship, i.e.
hypothesis formulation

Once the hypothesis is formulated the next
step is to test it

The analytic studies allow further analysis and
ultimately the rejection or acceptance of the
hypothesis
Observational
Case-control
Cohort

Experimental
Randomized control trial
Observations - cases
In 1940s Sir Norman Gregg, an Australian
ophthalmologist observed a number of infants and
young children with an unusual form of cataract.
Gregg noted that these children had been in utero
during the time of a rubella outbreak (German
measles). He suggested that there was an association
between prenatal rubella exposure and the
development of the unusual cataracts. At that time
there was no knowledge that a virus could be
teratogenic. Thus, he proposed his hypothesis solely
on the basis of observational data.

The investigator observes the occurrence of disease
in people who are already segregated into groups
on the basis of some experience or exposure
- e.g. married vs. single or smoker vs. non-smoker

In this type of study, allocation into groups on the
basis of exposure is not under the control of the
investigator

Contrasts between study groups are observed and
analyzed, not created experimentally
In an experiment the investigator studies
the impact of some varying factor that he
or she controls

The most common type of experimental
study encountered in medicine is the
randomized controlled trial

People diagnosed as having a disease
(cases) are compared with persons who do
not have the disease (controls)
The purpose is to determine if the two
groups differ by exposure
It compares cases and controls with regard
to the exposures in their past

Study which involves identifying patients who
have the outcome of interest (cases) and control
patients who do not have that same outcome,
and looking back to see if they had the
exposure of interest

The exposure could be some environmental
factor, a behavioral factor, or exposure to a
drug or other therapeutic intervention
Cases Controls
Exposed a b
Not exposed c d
TOTAL a+c b+d
Proportion exposed a/a+c b/b+d
CHD Cases Controls
Smoke cigarettes 112 176
Do not smoke
cigarettes
88 224
TOTAL 200 400
% Smoking cigarettes 56.0 44.0
Case-control Study design
Case-control Study design
Average daily
cigarettes
Lung cancer
patients
Control group
0 7 61
1-4 55 129
5-14 489 570
15-24 475 431
25-49 293 154
50+ 38 12
Total 1357 1357
Chronic diseases (e.g. cancer and CVD) have
long latency periods
Rare diseases (low population incidence) can
not be analyzed easily using another
approach
Time and money issues: cost efficient and
less time to complete
Many exposure factors can be studied at the
same time




Difficulties choosing appropriate controls
Can not get true estimate of risk
Issues with temporal association
Strong potential for bias
Confounding
I. Definition and selection of cases
II. Selection of controls
III. Ascertainment of disease and exposure
status
IV. Analysis
V. Interpretation of data-derived hypotheses
The definition must be specific
e.g. meningioma, not brain tumor
Use standardized diagnostic criteria
Selection of cases - representative of all cases
or a special group e.g. from a tertiary care
facility (easier to conduct) vs. community
based (more representative)


The selection of an appropriate control
group is the most difficult issue in case-
control design
They should be comparable to the source
population of the cases, including
exclusions and restrictions
No control group is optimal

Higher non-response rates
Less accurate medical records
Costly
May not be representative of the general
population (less healthy, less physically
active)
Less confounders


Community Controls
Neighborhood controls
Random-digit dialing controls
Best friend control
Spouse or sibling control


Easily identified and readily available
Medical records and health histories
available
Less non-response
Confounders: Hospitalized patients are
more likely to be smokers, alcoholics and
with other high risk behaviors
Issues: selecting a specific other
diagnosis as control
In 1981, MacMahon and coworkers reported a case-
control study of cancer of the pancreas. The cases were
patients with a histologically confirmed diagnosis of
pancreatic cancer in 11 Boston and Rhode Island
hospitals from 1974 to 1979. Controls were selected from
all patients who were hospitalized at the same time as
the cases; and they were selected from other inpatients
hospitalized by the attending physicians who had
hospitalized the cases. One finding in this study was an
apparent dose-response relationship between coffee
consumption and cancer of the pancreas, particularly in
women
Controls selection - issues
Case-control study of coffee drinking and
pancreatic cancer
Investigators usually use 2 or 3 control
but not more than 4 control groups,
selected in different ways, to minimize the
problems described
E.g Hurwitz, E. Public Health Study on
Reyes Syndrome NEJM 313:849, 1985
A case-control study to evaluate the association
between Reyes Syndrome and medication used
4 control groups emergency room, school,
community and hospitalized controls
Study revealed a strong association between
Reyes and aspirin use
Selection of Controls multiple controls
Gold and coworkers case-control study
of brain tumors in children
2 types of controls:
-children without cancer (normal controls)
-children with other than brain tumors
(cancer controls)
Selection of Controls multiple controls
Selection of Controls multiple controls
Disease:
death certificates, case-registries, office
records, hospital records, pathology log
books etc

Exposure:
Interview, mail questionnaires, medical
records etc



Bias is any systematic error in an
epidemiologic study that results in an
incorrect estimate of the risk of association
between exposure and risk of disease
All studies, but especially case-control
studies, have the potential for bias
The effects are difficult to evaluate in the
analysis
Bias should be eliminated, if possible,
when the study is designed
Selection bias
Recall bias
Interviewer bias
Experimenter expectancy
Misclassification bias
Measurement bias



= sampling bias
Sample selected is not representative of
the population
Sample selection may involve pre- or
post-selection that may preferentially
include or exclude certain kinds of
samples and therefore affecting the
results
Discounting trial subjects/tests that did not
run to completion. For example, in a test of a
dieting program, the researcher may simply
reject everyone who drops out of the trial.
But most of those who drop out are those for
whom it wasn't working.

A CC study to predict rate of heart disease
in a community and using members of
health club for seniors as sample






Recall bias occurs whenever individuals with a
particular adverse effect outcome remember their
previous exposure experience differently from those
who arent similarly affected
People who are sick tend to think about possible
causes for their illness
Recall bias can lead to an over or underestimation
of risk




A study of prenatal infections and congenital
malformations:
-cases mothers of children with congenital
malformations
-controls mothers of children without
congenital malformations

Recall Bias - example
Cases (congenital
malformations)
Controls (without
congenital
malformations)
Assume that:
True incidence infection (%) 15 15
Infections recalled (%) 60 10
Interviewers who are aware of the study
hypothesis are likely to question cases and
controls differently

More probing questions may be asked of cases

The interviewer may unconsciously sabotage the
process

Pygmalion effect

Experimenters expectations are
communicated to the subjects,
unintentionally

The subject then produce the desired
effect


Misclassification: Subjects may be
erroneously categorized with respect to
exposure or disease status
Measurement: Method of collecting
information was flawed and asked leading
questions



False estimate of survival rates

A screening test identifies a disease much
before its usual detection (clinical sign),
thus make the survival seem longer
Selection bias: controls picked from same
source as cases, use motivated
individuals
Recall bias: one hospitalized control
group
Interviewer bias: highly trained
personnel, blinded to study hypothesis,
standardized questionnaires etc
Misclassification bias: use more than one
source to validate

A potential confounder is the variable that is
known to be associated with the outcome
(effect) even though it is not the variable under
study
Exposure Disease
Confounder
Common confounders include age, gender,
tobacco, alcohol, socio-economic status
e.g. A case-control study shows an
association between decreased level of
physical activity and increased risk of MI

Could age be a confounder?
Study design:
Randomization
Matching
Restriction
Analysis:
Stratified analysis
Multivariate analysis
Cases and controls are matched by usual
confounders (e.g. age, sex, SES, smoking,
alcohol etc.) so that these factors are
equally distributed in both groups and will
not confound the association between the
variables
Can be: group matching or individual
matching
Disadvantage: It can be very difficult and
expensive to find a perfect match for each
case
Another way to reduce the effect of
confounders in a study is to place restrictions
on the study subjects
If smoking can be a confounder, then only
enroll non-smokers
If age is a confounder then place age
restrictions
Disadvantage: Too many restrictions can lead
to difficulty with generalizability

Create strata of the confounding variable
If sex is a confounder then analyze men and
women separately
If age is a confounder then analyze data
separately for each age group
Disadvantage:
It is extremely cumbersome
Difficult to control for more than 1
confounder at a time

V. Case-control Study and the
Odds Ratio

Incidence can not be derived in a case-
control study

The estimate of relative risk (odds
ratio) can be calculated
ODDS RATIO
Example
Of 200 patients in the hospital, 50 have lung
cancer. Of these 50 patients, 45 are smokers. Of
the remaining 150 hospitalized patients who do
not have lung cancer, 60 are smokers. This
information can be used to calculate the odds
ratio for smoking and the risk of lung cancer.



ODDS RATIO
with LC without LC
smokers A=45 B=60
nonsmokers C=5 D=90

Odds Ratio =
ratio of odds the cases that were exposed
ratio of odds the controls that were exposed

OR = A/C or (A)(D) = (45)(90) = 13.5
B/D (B)(C) (5)(60)


Analysis of Case-control Studies
a
d
b
Case Control
Risk factor
Yes



No
Odds Ratio = ad
bc
c
Odds Ratio (OR)
A ratio that measures the odds of exposure for cases
compared to odds of exposure for controls

Odds of exposure = no. exposed no. unexposed

OR Numerator: Odds of exposure for cases

OR Denominator: Odds of exposure for controls

Interpreting the Odds Ratio
The odds of exposure for cases are
1.62 times the odds of exposure for
controls.

Those with CHD are 1.62 times more likely to be
smokers than those without CHD

CLINICAL PROBABILITY

The probability of an event can be
expressed as a ratio of the number of
likely outcomes to the number of possible
outcomes

The probability of an event is denoted by P
Probabilities are usually expressed as
decimals fractions, not as percentages, and
must lie between zero (zero probability)
and one (absolute certainty)





Probability
Methods of calculating probability:

The multiplication rule
The addition rule

Multiplication rule
The multiplication rule of probability
states that the probability of two or more
statistically independent events all-
occurring at same time is equal to the
product of their individual probabilities
Example:
Chance of having a brown hair is 0.3
Chance of getting a cold is 0.2
What is the chance of meeting brown
haired person with a cold?
0.30.2 = 0.06

Addition rule
The addition rule of probability states that the
probability of any one of several particular
events occurring is equal to the sum of their
individual probabilities, provided the events
are mutually exclusive (i.e. they cannot happen
at one time)
Example:
The probability of picking a heart card in a deck is
0.25, The probability of picking a diamond card
in a deck is 0.25. what is the probability of
picking a heart or a diamond card?

0.25+0.25 = 0.5

Was the study design appropriate?
How were cases diagnosed and selected
How were controls selected - appropriate?
How was bias minimized?
How was confounding controlled?
Did the authors identify areas of bias and
confounding?