Biases and errors in

Epidemiology
Anchita Khatri
 Definitions
ERROR:
1. A false or mistaken result obtained in a study
or experiment
2. Random error is the portion of variation in
measurement that has no apparent connection
to any other measurement or variable,
generally regarded as due to chance
3. Systematic error which often has a
recognizable source, e.g., a faulty measuring
instrument, or pattern, e.g., it is consistently
wrong in a particular direction
(Last)
 Relationship b/w Bias and Chance
True BP BP measurement
(intra-arterial (sphygmomanometer)
cannula)
No. of observations

Chance

Bias
80 90
Diastolic Blood Pressure (mm Hg)
 Validity
Validity: The degree to which a
measurement measures what it purports to
measure (Last)
Degree to which the data measure what they
were intended to measure that is, the
results of a measurement correspond to the
true state of the phenomenon being
measured (Fletcher)
also known as Accuracy
 Reliability
The degree of stability expected when a
measurement is repeated under identical conditions;
degree to which the results obtained from a
measurement procedure can be replicated
(Last)
Extent to which repeated measurements of a stable
phenomenon by different people and instruments,
at different times and places get similar results
(Fletcher)
Also known as Reproduciblity and Precision
 Validity and Reliability

VALIDITY

High Low

High
RELIABILITY

Low
 Bias
Deviation of results or inferences from the truth,
or processes leading to such deviation. Any
trend in the collection, analysis, interpretation,
publication, or review of data that can lead to
conclusions that are systematically different
from the truth. (Last)
A process at any stage of inference tending to
produce results that depart systematically from
true values (Fletcher)
 Types of biases

1. Selection bias
2. Measurement / (mis)classification bias
3. Confounding bias
 Selection bias
Errors due to systematic differences in
characteristics between those who are selected
for study and those who are not.
(Last; Beaglehole)
When comparisons are made between groups
of patients that differ in ways other than the
main factors under study, that affect the
outcome under study. (Fletcher)
 Examples of Selection bias
Subjects: hospital cases under the care of a
physician
Excluded:
1. Die before admission acute/severe disease.
2. Not sick enough to require hospital care
3. Do not have access due to cost, distance etc.
Result: conclusions cannot be generalized
Also known as Ascertainment Bias
(Last)
 Ascertainment Bias
Systematic failure to represent equally all
classes of cases or persons supposed to be
represented in a sample. This bias may arise
because of the nature of the sources from
which the persons come, e.g., a specialized
clinic; from a diagnostic process influenced
by culture, custom, or idiosyncracy . (Last)
 Selection bias with volunteers
Also known as response bias
Systematic error due to differences in
characteristics b/w those who choose or
volunteer to take part in a study and those
who do not
 Examples response bias

Volunteer either because they are unwell, or

worried about an exposure
Respondents to effects of smoking usually not
as heavy smokers as non-respondents.
In a cohort study of newborn children, the
proportion successfully followed up for 12
months varied according to the income level of
the parents
 Examples. (Assembly bias)
Study: ? association b/w reserpine and breast
cancer in women
Design: Case Control
Cases: Women with breast cancer
Controls: Women without breast cancer
who were not suffering from any
cardio-vascular disease (frequently
associated with HT)
Result: Controls likely to be on reserpine
systematically excluded association between
reserpine and breast cancer observed
 Examples. (Assembly bias)
Study: effectiveness of OCP1 vs. OCP2
Subjects:
on OCP1 women who had given birth at least
once ( able to conceive)
on OCP2 women had never become pregnant
Result: if OCP2 found to be better, inference
correct??
 Susceptibility Bias

Groups being compared are not equally

susceptible to the outcome of interest, for
reasons other than the factors under study
Comparable to Assembly Bias
In prognosis studies; cohorts may differ in
one or more ways extent of disease,
presence of other diseases, the point of time
in the course of disease, prior treatment etc.
 Examples..(Susceptibility Bias)
Background: for colorectal cancer,
- CEA levels correlated with extent of disease
(Dukes classification)
- Dukes classification and CEA levels strongly
predicted diseases relapse
Question: Does CEA level predict relapse
independent of of Dukes classification, or was
susceptibility to relapse explained by Dukes
classification alone?
 Example CEA levels (contd.)
Answer: association of pre-op levels of
CEA to disease relapse was observed for
each category of Dukes classification

stratification
 Disease-free survival according to CEA
levels in colorectal cancer pts.with similar
pathological staging (Dukes B)
100
% disease free

80
CEA Level (ng)
<2.5
2.5 10.0
60
>10.0

0 3 6 9 12 15 18 21 24
Months
Selection bias with Survival Cohorts
Patients are included in study because they are
available, and currently have the disease
For lethal diseases patients in survival cohort are
the ones who are fortunate to have survived, and
so are available for observation
For remitting diseases patients are those who are
unfortunate enough to have persistent disease
Also known as Available patient cohorts
Example bias with survival cohort
TRUE COHORT Observed True
Measure outcome improvement improvement
Assemble Improved: 75
Cohort
(N=150)
Not improved: 75 50% 50%

SURVIVAL COHORT
Assemble
patients
Begin Measure outcome 50%
Follow-up 80%
Improved: 40
(N=50) Not improved: 10
Not observed Dropouts
(N=100) Improved: 35
Not improved: 65
 Selection bias due to Loss to
Follow-up
Also known as Migration Bias
In nearly all large studies some members of the
original cohort drop out of the study
If drop-outs occur randomly, such that
characteristics of lost subjects in one group are
on an average similar to those who remain in
the group, no bias is introduced
But ordinarily the characteristics of the lost
subjects are not the same
 Example of lost to follow-up
EXPOSURE EXPOSURE
irradiation irradiation
+nt -nt Total +nt -nt Total
DISEASE
cataract

+nt 50 100 150 +nt 30 30 60

-nt -nt
10000 20000 30000 4000 8000 12000
RR= 50/10000 RR= 30/4000
100/20000 30/8000
=1 =2
 Migration bias
A form of Selection Bias
Can occur when patients in one group leave
their original group, dropping out of the
study altogether or moving to one of the
other groups under study (Fletcher)
If occur on a large scale, can affect validity
of conclusions.
Bias due to crossover more often a problem
in risk studies, than in prognosis studies,
because risk studies go on for many years
 Example of migration
Question: relationship between lifestyle and
mortality
Subjects: 10,269 Harvard College alumni
- classified according to physical activity,
smoking, weight, BP
- In 1966 and 1977
Mortality rates observed from 1977 to 1985
 Example of migration (contd.)
Problem: original classification of lifestyle
might change (migration b/w groups)
Solution: defined four categories
- Men who maintained high risk lifestyles
- Men who crossed over from low to high risk
- Men who crossed over from high to low risk
- Men who maintained low risk lifestyles
 Example of migration (contd.)
Result: after controlling for other risk
factors
- those who maintained or adopted high risk
characteristics had highest mortality
- Those who changed from high to low had
lesser mortality than above
- Those who never had any high risk
behavior had least mortality
 Healthy worker effect
A phenomenon observed initially in studies of
occupational diseases: workers usually
exhibit lower overall death rates than the
general population, because the severely ill
and chronically disabled are ordinarily
excluded from employment. Death rates in
the general population may be inappropriate
for comparison if this effect is not taken into
account.
(Last)
Example. healthy worker effect

Question: association b/w formaldehyde

exposure and eye irritation
Subjects: factory workers exposed to
formaldehyde
Bias: those who suffer most from eye
irritation are likely to leave the job at their
own request or on medical advice
Result: remaining workers are less affected;
association effect is diluted
 Measurement bias
Systematic error arising from inaccurate
measurements (or classification) of subjects or
study variables. (Last)
Occurs when individual measurements or
classifications of disease or exposure are
inaccurate (i.e. they do not measure correctly
what they are supposed to measure)
(Beaglehole)
If patients in one group stand a better chance of
having their outcomes detected than those in
another group. (Fletcher)
Measurement / (Mis) classification

Exposure misclassification occurs when

exposed subjects are incorrectly classified
as unexposed, or vice versa
Disease misclassification occurs when
diseased subjects are incorrectly classified
as non-diseased, or vice versa
(Norell)
 Causes of misclassification
1. Measurement gap: gap between the
measured and the true value of a variable
- Observer / interviewer bias
- Recall bias
- Reporting bias
2. Gap b/w the theoretical and empirical
definition of exposure / disease
Sources of misclassification

Measurement results

Measurement errors

Empirical definition

Gap b/w theoretical & empirical definitions

Theoretical definition
 Example gap b/w definitions
Theoretical definition Empirical definition
Exposure: passive Exposure: passive
smoking inhalation of smoking time spent with
tobacco smoke from smokers (having smokers
as room-mates)
other peoples smoking
Disease: Myocardial
Disease: Myocardial infarction certain
infarction necrosis of diagnostic criteria (chest
the heart muscle tissue pain, enzyme levels, signs
on ECG)
 Exposure misclassification
Non-differential
Misclassification does not differ between
cases and non-cases
Generally leads to dilution of effect, i.e.
bias towards RR=1 (no association)
 ExampleNon-differential
Exposure Misclassification

EXPOSURE EXPOSURE
X-ray exposure X-ray exposure
+nt -nt Total +nt -nt Total
Breast Cancer
DISEASE

+nt 40 80 120 +nt 60 60 120

-nt -nt
10000 40000 50000 20000 30000 50000
RR= 40/10000 RR= 60/20000
80/40000 60/30000
=2 = 1.5
 Exposure misclassification -
Differential
Misclassification differs between cases and
non-cases
Introduces a bias towards
RR= 0 (negative / protective association), or
RR= (infinity)(strong positive association)
 ExampleDifferential Exposure
Misclassification
EXPOSURE EXPOSURE
X-ray exposure X-ray exposure
+nt -nt Total +nt -nt Total
Breast Cancer
DISEASE

+nt 40 80 120 +nt 40 80 120

-nt 9960 39920 49880 -nt 19940 29940 49880
10000 40000 50000 19980 30020 50000
RR= 40/10000 RR= 40/19980
80/40000 80/30020
=2 = 0.75
 Implications of Differential
exposure misclassification
An improvement in accuracy of exposure
information (i.e. no misclassification among
those who had breast cancer), actually
reduced accuracy of results
Non-differential misclassification is better
than differential misclassification
So, epidemiologists are more concerned
with comparability of information than
with improving accuracy of information
 Causes of Differential Exposure
Misclassification
Recall Bias:Systematic error due to
differences in accuracy or completeness of
recall to memory of past events or
experience.
For e.g. patients suffering from MI are more
likely to recall and report lack of exercise
in the past than controls
 Causes of Differential Exposure
Misclassification
Measurement bias:
e.g. analysis of Hb by different methods
(cyanmethemoglobin and Sahli's) in cases
and controls.
e.g.biochemical analysis of the two groups
from two different laboratories, which give
consistently different results
 Causes of Differential Exposure
Misclassification
Interviewer / observer bias: systematic error
due to observer variation (failure of the
observer to measure or identify a
phenomenon correctly)
e.g. in patients of thrombo-embolism, look for
h/o OCP use more aggressively
 Measurement bias in treatment
effects
Hawthorne effect: effect (usually positive /
beneficial) of being under study upon the
persons being studied; their knowledge of
being studied influences their behavior
Placebo effect: (usually, but not necessarily
beneficial) expectation that regimen will
have effect, i.e. the effect is due to the
power of suggestion.
 Total effects of treatment are the sum of
spontaneous improvement, non-specific responses,
and the effects of specific treatments
EFFECTS
Specific to
treatment
IMPROVEMENT

Placebo

Hawthorne

Natural
History
 Confounding
1. A situation in which the effects of two
processes are not separated. The distortion
of the apparent effect of an exposure on risk
brought about by the association with other
factors that can influence the outcome
2. A relationship b/w the effects of two or
more causal factors as observed in a set of
data such that it is not logically possible to
separate the contribution that any single
causal factor has made to an effect
(Last)
 Confounding
When another exposure exists in the study
population (besides the one being studied)
and is associated both with disease and the
exposure being studied. If this extraneous
factor itself a determinant of or risk factor
for health outcome is unequally distributed
b/w the exposure subgroups, it can lead to
confounding
(Beaglehole)
 Confounder must be
1. Risk factor among the unexposed (itself a
determinant of disease)
2. Associated with the exposure under study
3. Unequally distributed among the exposed
and the unexposed groups
 Examples confounding

SMOKING LUNG CANCER

(As age advances

chances of lung
cancer increase)
AGE
(If the average ages of the smoking and
non-smoking groups are very different)
 Examples confounding

COFFEE DRINKING HEART DISEASE

(Smoking increases
(Coffee drinkers are the risk of heart ds)
more likely to smoke)

SMOKING
 Examples confounding

ALCOHOL MYOCARDIAL
INTAKE INFARCTION

(Men are more likely (Men are more at risk

to consume alcohol for MI)
than women)

SEX
 Examples confounding
Exposure-alcohol

+nt -nt
RR = 140/30000
+nt 140 100
Disease

100/30000
MI

-nt = 1.4
Total 30000 30000
Exposure-alcohol
+nt -nt RR = 120/20000
(M) 60/10000
male female male female
=1
+nt 120 20 60 40 RR = 20/10000
Disease
MI

-nt (F) 40/20000

Total 20000 10000 10000 20000 =1
 Example multiple biases
Study: ?? Association b/w regular exercise and
risk of CHD
Methodology: employees of a plant offered an
exercise program; some volunteered, others did
not
coronary events detected by regular voluntary
check-ups, including a careful history, ECG,
checking routine heath records
Result: the group that exercised had lower CHD
rates
 Biases operating
Selection: volunteers might have had initial
lower risk (e.g. lower lipids etc.)
Measurement: exercise group had a better
chance of having a coronary event detected
since more likely to be examined more
frequently
Confounding: if exercise group smoked
cigarettes less, a known risk factor for CHD
 Dealing with Selection Bias
Ideally,
To judge the effect of an exposure / factor on
the risk / prognosis of disease, we should
compare groups with and without that
factor, everything else being equal
But in real life everything else is usually not
equal
Methods for controlling Selection Bias
During Study Design
1. Randomization
2. Restriction
3. Matching
During analysis
1. Stratification
2. Adjustment
a) Simple / standardization
b) Multiple / multivariate adjustment
c) Best case / worst case analysis
 Restriction
Subjects chosen for study are restricted to
only those possessing a narrow range of
characteristics, to equalize important
extraneous factors
Limitation: generalisability is
compromised; by excluding potential
subjects, cohorts / groups selected may be
unusual and not representative of most
patients or people with condition
 Example restriction
Study: effect of age on prognosis of MI
Restriction: Male / White / Uncomplicated
anterior wall MI
Important extraneous factors controlled for:
sex / race / severity of disease
Limitation: results not generalizable to
females, people of non-white community,
those with complicated MI
 Example restriction
OCP example
restrict study to women having at least one
child

Colorectal cancer example

restrict patients to a particular staging of
Dukes classification
 Matching - definition
The process of making a study group and a
comparison group comparable with respect to
extraneous factors (Last)
For each patient in one group there are one or
more patients in the comparison group with
same characteristics, except for the factor of
interest (Fletcher)
 Types of Matching
Caliper matching: process of matching
comparison group to study group within a
specific distance for a continuous variable
(e.g., matching age to within 2 years)
Frequency matching: frequency distributions of
the matched variable(s) be similar in study and
comparison groups
Category matching: matching the groups in
broad classes such as relatively wide age
ranges or occupational groups
 Types of Matching (contd.)
Individual matching: identifying individual
subjects for comparison, each resembling a
study subject on the matched variable(s)
Pair matching: individual matching in
which the study and comparison subjects
are paired

(Last)
 Matching is often done for age, sex, race, place
of residence, severity of disease, rate of
progression of disease, previous treatment
received etc.
Limitations:
- controls for bias for only those factors involved
in the match
- Usually not possible to match for more than a
few factors because of the practical difficulties
of finding patients that meet all matching criteria
- If categories for matching are relatively crude,
there may be room for substantial differences
b/w matched groups
 Example Matching
Study: ? Association of Sickle cell trait (HbAS)
with defects in physical growth and cognitive
development
Other potential biasing factors: race, sex, birth
date, birth weight, gestational age, 5-min Apgar
score, socio economic status
Solution: matching for each child with HbAS
selected a child with HbAA who was similar with
respect to the seven other factors (50+50=100)
Result: no difference in growth and development
 Overmatching
A situation that may arise when groups are being
matched. Several varieties:
1. The matching procedure partially or
completely obscures evidence of a true causal
association b/w the independent and
dependant variables. Overmatching may occur
if the matching variable is involved in, or is
closely connected with, the mechanism
whereby the independent variable affects the
dependant variable. The matching variable
may be an intermediate cause in the causal
chain or it may be strongly affected by, or a
consequence of, such an intermediate cause
2. The matching procedure uses one or more
unnecessary matching variables, e.g., variables
that have no causal effect or influence on the
dependant variable, and hence cannot confound
the relationship b/w the independent and
dependant variables.
3. The matching process is unduly elaborate,
involving the use of numerous matching
variables and / or insisting on a very close
similarity with respect to specific matching
variables. This leads to difficulty in finding
suitable controls (Last)
 Stratification
The process of or the result of separating a
sample into several sub-samples according
to specified criteria such as age groups,
socio-economic status etc. (Last)
The effect of confounding variables may be
controlled by stratifying the analysis of
results
After data are collected, they can be
analyzed and results presented according to
subgroups of patients, or strata, of similar
characteristics (Fletcher)
ExampleStratification (Fletcher)

HOSPITAL A
Pre-op
Risk
Pts Deaths %
High 500 30 6
Medium 400 16 4
HOSPITAL B
Low 300 02 .67
Pre-op
Total 1200 48 4 Risk
Pts Deaths %
High 400 24 6
Medium 800 32 4
Low 1200 8 .67
Total 2400 64 2.6
 ExampleStratification
Age Relat.
Pinellas county Dade county
Wise Rate
Stratifi
cation Dead Total Rate Dead Total Rate

Birth 737
229,198 3.2 2463 748,035 3.3 1.0
54 yrs

> 55 yrs 4989 145,147 34.4 5898 187,985 31.2 1.1

Overall 5726 374,665 15.3 8332 935,047 8.9 1.7

 Standardization

A set of techniques used to remove as far as

possible the effects of differences in age or other
confounding variables when comparing two or
more populations
The method uses weighted averaging of rates
specific for age, sex, or some other potentially
confounding variable(s), according to some
specified distribution of these variables
(Last)
 Standard population
A population in which the age and sex
composition is known precisely, as a result
of a census or by an arbitrary means e.g.
an imaginary population, the standard
million in which the age and sex
composition is arbitrary. A standard
population is used as comparison group in
the actuarial procedure of standardization of
mortality rates. (e.g. Segi world population,
European standard population)
(Last)
 Types of standardization

Direct: the specific rates in a study population

are averaged using as weights the
distribution of a specified standard
population.
The standardized rate so obtained represents
what the rate would have been in the study
population if that population had the same
distribution as the standard population w.r.t.
the variables for which the adjustment or
standardization was carried out.
Indirect: used to compare the study populations
for which the specific rates are either
statistically unstable or unknown. The specific
rates are averaged using as weights the
distribution of the study population. The ratio
of the crude rate for the study population to
the weighted average so obtained is known as
standardized mortality (or morbidity) ratio, or
SMR. (Last)
[represents what the rate would have been in the
study population if that population had the
same specific rates as the standard population]
 Standardized mortality ratio
(SMR)
Ratio of
The no. of deaths observed in the
study group or population
X 100
No. of deaths expected if the study
population had the same specific rates
as the standard population
 Example direct standardization
Age Pop Deaths Rate Std.Pop Exp deaths
0 4000 60 15.0 2400 36
1-4 4500 20 4.4 9600 42.24
5-14 4000 12 3.0 19000 57
15-19 5000 15 3.0 9000 27
20-24 4000 16 4.0 8000 32
25-34 8000 25 3.1 14000 43.4
34-44 9000 48 5.3 12000 63.6
45-54 8000 100 12.5 11000 137.5
55-64 7000 150 21.4 8000 171.2
Total 53,500 446 8.3 93000 609.94(6.56)
 Example direct standardization
HOSPITAL A
Preop Pts Deaths %
High 500 30 6
Medium 400 16 4 HOSPITAL Std
Low 300 2 .67 Preop Pts Rate Exp.deaths
Total 1200 48 4 High 400 6 24
Medium 400 4 16
Low 400 .67 2.68
Total 1200 42.68
(3.6%)
 Stratification vs. Standardization
Standardization removes the effect
Stratification controls for the effect of factor,
but the effect can still be seen
For e.g. in the hospital example, with
standardization we found that patients had
similar prognosis in both hospitals; with
stratification also learnt mortality rates among
different risk strata
Similar to difference b/w age-standardized
mortality rate and age specific mortality rates
 Multivariate adjustment
Simultaneously controlling the effects of
many variables to determine the independent
effects of one
Can select from a large no. of variables a
smaller subset that independently and
significantly contributes to the overall
variation in outcome, and can arrange
variables in order of the strength of their
contribution
Only feasible way to deal with many variables
at one time during the analysis phase
 Examples Multivariate
adjustment
CHD is the joint result of lipid
abnormalities, HT, smoking, family history,
DM, exercise, personality type.
Start with 2x2 tables using one variable at a
time
Contingency tables, i.e. stratified analyses,
examining the effect of one variable
changed in the presence/absence of one or
more variables
ExampleMultivariate adjustment
Multi variable modeling i.e developing a
mathematical expression of the effects of many
variables taken together
Basic structure of a multivariate model:
Outcome variable = constant + (1 x variable1)
+ (2 x variable2) + .
1, 2, are coefficients determined from the
data; variable1, variable2, . are the predictor
variables that might be related to outcome
 Sensitivity analysis
When data on important prognostic factors
is not available, it is possible to estimate the
potential effects on the study by assuming
various degrees of mal-distribution of the
factors b/w the groups being compared and
seeing how that would affect the results
Best case / worst case analysis is a special
type of sensitivity analysis assuming the
best and worst type of mal-distribution
Example best/worst case analysis

Study: effect of gastro-gastrostomy on

morbid obesity
Subjects: cohort of 123 morbidly obese
patients who underwent gastro-gastrostomy,
19 to 47 months after surgery
Success : losing >30% excess weight
Follow-up: 103 (84%) patients
20 patients lost to follow up
 Example. (contd.)
Success rate: 60/103 (58%)
Best case: all 20 lost to follow up had
success
Best success rate: (60+20)/123 (65%)
Worst case: all 20 lost to follow up had
failures
Worst success rate: 60/123 (49%)
Result: true success rate b/w 49% and 65%;
probably closer to 58% ! (because pts. lost to
follow up unlikely to be all successes or all
failures
 Randomization
The only way to equalize all extraneous
factors, or everything else is to assign
patients to groups randomly so that each has
an equal chance of falling into the exposed
or unexposed group
Equalizes even those factors which we
might not know about!
But it is not possible always
 Overall strategy
Except for randomization, all ways of
dealing with extraneous differences b/w
groups. Are effective against only those
factors that are singled out for consideration
Ordinarily one uses several methods layered
one upon another
 Example
Study: effect of presence of VPCs on survival of
patients after acute MI
Strategies:
- Restriction: not too young / old; no unusual
causes (e.g.mycotic aneurysm) for infarction
- Matching: for age (as important prognostic
factor, but not the factor under study)
- Stratification: examine results for different
strata of clinical severity
- Multivariate analysis: adjust crude rates for the
effects of all other variables except VPC, taken
together.
 Dealing with measurement bias
1. Blinding
- Subject
- Observer / interviewer
- Analyser
2. Strict definition / standard definition for
exposure / disease / outcome
3. Equal efforts to discover events equally in
all the groups
 Controlling confounding
Similar to controlling for selection bias
Use randomization, restriction, matching,
stratification, standardization, multivariate
analysis etc.
 Lead time bias
Lead time is the period of time b/w the
detection of a medical condition by screening
and when it ordinarily would be diagnosed
because a pt. experiences symptoms and seeks
medical care
As a result of screening, on an average, pt will
survive longer from the time of diagnosis than
who are diagnosed otherwise, even if T/t is
not effective.
Not more survival time, but more disease
time
 How lead time affects survival time

Unscreened Diag

Screened Diag
Early T/t not effective

Screened Diag
Early T/t is effective

Onset of Ds Death Survival after

diagnosis
 Controlling lead time bias
Compare screened group of people, and
control group, and compare age specific
mortality rates, rather than survival times
from time of diagnoses
E.g. early diagnosis and T/t for colorectal
cancer is effective because mortality rates of
screened people are lower than those of a
comparable group of unscreened people
 Length time bias
Can affect studies of screening
Bcos the proportion of slow growing tumors
diagnosed during screening programs is
greater than those diagnosed during usual
medical care
Bcos slow growing tumors are present for a
longer period before they cause symptoms;
fast growing tumors are likely to cause
symptoms leading to interval diagnosis
Screening tends to find tumors with
inherently better prognoses
 Compliance bias
Compliant patients tend to have better
prognoses regardless of the screening
If a study compares disease outcomes
among volunteers for a screening program
with outcomes in a group of people who did
not volunteer, better results for the
volunteers might not be due to T/t but due
to factors related to compliance
Compliance bias and length-time bias can
both be avoided by relying on RCTs
 Types of studies & related biases
Prevalence study Uncertainty about temporal sequences
Bias studying old/prevalent cases
Case control Selection bias in selecting
cases/controls
Measurement bias
Cohort study Susceptibility bias
Survival cohort vs. true cohort
Migration bias
Randomized Consider natural h/o disease,
control trials Hawthorne effect, placebo effect etc.
Compliance problems
Effect of co-interventions
 Random error
Divergence on the basis of chance alone of
an observation on a sample from the
population from the true population values
random because on an average it is as
likely to result in observed values being on
one side of the true value as on the other
side
Inherent in all observations
Can be minimized, but never avoided
altogether
 Sources of random error
1. Individual biological variation
2. Measurement error
3. Sampling error ( the part of the total
estimation of error of a parameter caused
by the random nature of the sample)
 Sampling variation
Because research must ordinarily be
conducted on a sample of patients and not
on all the patients with the condition under
study

always a possibility that the particular

sample of patients in a study, even though
selected in an unbiased way, might not be
similar to population of patients as a whole
Sampling variation - definition
Since inclusion of individuals in a sample is
determined by chance, the results of
analysis on two or more samples will differ
purely by chance.
(Last)
 Assessing the role of chance

1. Hypothesis testing
2. Estimation
 Hypothesis testing
Start off with the Null Hypothesis (H0)
the statistical hypothesis that one variable
has no association with another variable or
set of variables, or that two or more
population distributions do not differ from
one another.
in simpler terms, the null hypothesis states
that the results observed in a study,
experiment or test are no different from
what might have occurred as a result of
operation of chance alone
 Statistical tests errors
(Fletcher)

TRUE DIFFERENCE

PRESENT ABSENT
(H0) false (H0) true

SIGNIFICANT Type I
CONCLUSION (H ) Rejected
0
Power ( ) error
OF
STATISTICAL NOT
TEST SIGNIFICANT Type II
(H0) Accepted ( ) error
 Statistical tests - errors
Type I () error: error of rejecting a true
null hypothesis , I.e. declaring a difference
exists when it does not
Type II () error: error of failing to reject a
false null hypothesis , I.e. declaring that a
difference does not exist when in fact it
does
Power of a study: ability of a study to
demonstrate an association if one exists
Power = 1-
 p - value
Probability of an error.
Quantitative estimate of probability that
observed difference in b/w the groups in the
study could have happened by chance alone,
assuming that there is no real difference b/w the
groups OR
If there were no difference b/w the groups, and
the trial was repeated many times, what
proportion of the trials would lead to conclusions
that there is the same or a bigger difference b/w
the groups than the results found in the study
 p value Remember!!
Usually P < 0.05 is considered statistically
significant (i.e. probability of 1 in 20 that
observed difference is due to chance)
0.05 is an arbitrary cut-off; can change
according to requirements
Statistically significant result might not be
clinically significant and vice-versa
 Statistical significance vs.
clinical significance
Large RCT called GUSTO (41,021 pts of ac MI)
Study: Streptokinase vs. tPA
Result: death rate at 30 days
- streptokinase (7.2%) (p < 0.001)
- tPA (6.3%)
But, need to treat 100 patients with tPA instead
of streptokinase to prevent 1 death!
tPA costly - $ 250 thousand to save one death
??? Clinically significant
 Estimation
Effect size observed in a particular study is
called Point estimate
True effect is unlikely to be exactly that
observed in study because of random
variation
Confidence interval (CI): usually 95%
(Last) computed interval with a given
probability e.g. 95%, that the true value
such as a mean, proportion, or rate is
contained within the interval
 Confidence intervals
(Fletcher) If the study is unbiased, there is a 95%
chance that that the interval includes the true
effect size. The true value is likely to be close
to the point estimate, less likely to be near the
outer limits of that interval, and could (5 times
out of 100) fall outside these limits altogether,

CI allows the reader to see the range of

plausible values and so to decide whether the
effect size they regard as clinically meaningful
is consistent with or ruled out by the data
 Multiple comparison problem

If a no. of comparisons are made, (e.g. in a

large study, the effect of treatment assessed
separately for each subgroup, and for each
outcome), 1 in 20 of these comparisons is
likely to be statistically significant at the
0.05 level
If you dredge the data sufficiently deeply,
and sufficiently often, you will find
something odd. Many of these bizarre
findings will be due to
chance.discoveries that were not
initially postulated among the major
objectives of the trial should be treated with
extreme caution.
 Dealing with random error
Increasing the sample size: sample size
depends upon
- level of statistical significance ( error)
- Acceptable chance of missing a real effect (
error)
- Magnitude of effect under investigation
- Amount of disease in population
- Relative sizes of groups being compared
Sample size is usually a compromise b/w
ideal and logistic and financial considerations
 References
1. Fletcher RH et al.Clinical Epidemiology : The
Essentials 3rd ed.
2. Beaglehole R et al. Basic Epidemiology, WHO
3. Last JM. Dictionary in Epidemiology 3rd ed.
4. Maxcy-Rosenau-Last. Public Health &
Preventive Medicine 14th ed.
5. Norell SE. Workbook of Epidemiology
6. Park K. Parks textbook of preventive and social
medicine 16th ed.