# Bias by liuqingyan

VIEWS: 17 PAGES: 37

• pg 1
```									Epidemiology

Lecture 7 : Sources of Error:
Selection and Measurement Bias
Learning outcomes

 Identify major sources of bias in
research studies
 Describe ways to overcome bias in
research studies
research
 Aim of any study is to estimate the
„true value‟ of a health event/disease
or association
 The best that can be achieved in the
real world is to make an empirical
estimate of the true value (theoretical
value) by observing the experience of
a sample of the population that is
specific in time and place
Possible outcomes of the study

Results of the           Truth – Actual
Study              No effect          Effect

No effect           Correct       Error (type 2)

Effect           Error (type 1)      Correct

 True value ~ is known as theoretical value
 Need to identify and control sources of
error
Validity
 External validity - extent to which the results
of a study can apply to people outside the
study – the target population
ie how generalisable are the results to others
 Internal validity - extent to which a study
properly measures what it is meant to:
 ie whether the conclusions are 'true'
for the people in the study.
Validity
Internal Validity

Target          Study       Exposed
Population                   Outcomes   Results
Population      or Sample    Unexposed

External Validity
research
 Error means a departure of the
empirical (observed) estimate from its
theoretical value „true value‟

Errors can occur:
 By Chance  Random Errors
 Systematically     Systematic Errors
Bias
Sources of Error

 Random Error
 Selection Bias
 Information Bias
 Confounding
Random Error (Random
Sampling Error)
 Difference between an empirical
(observed) estimate and its true
value that is due to chance variation
Example: Prevalence of arthritis
in women 45-69years
 Study 1: a survey - 1 in 20 sample of women
aged 45 to 69 years resident in Lidcombe in
May 1994
   Results: Arthritis Prevalence: 30 /100 population
.
 A survey using another 1 in 20 random
sample of women aged 45 to 69 years
   Results: Arthritis Prevalence: 40/100 population
 Entire population was surveyed
   Results: Arthritis Prevalence: 38/100 population
Random sampling variability

 The random sample surveys differ
from the true population value
because of random sampling
variability
Random sampling variability

 How confident can we be that our
estimate is close to the true value?
 A confidence interval is a range
within which the true value is
expected to lie
 Large samples have smaller CIs as
they are less prone to chance error
Bias
 „systematic distortion or deviation in the
study results from the true value‟
 „systematic difference between what the
study estimates and what it is intended to
estimate‟
 Systematic error (bias) is present when
the theoretical value of an epidemiological
measure cannot be obtained from a study
even in the total absence of random error
  overestimation or underestimation of
the effect due to a deficiency in the design
or execution of the study
Bias
  the likelihood that the study results will
be distorted from the true value
 Bias results from systematic flaws in
   study design
   data collection
   analysis of results
Example - Prevalence of arthritis

Lidcombe study and failed to include
residents from two large nursing
homes.
Result: arthritis prevalence:25/100
 Our prevalence estimate has a
systematic error or bias
Detecting Bias
 Usually impossible to detect bias by
looking at a prevalence estimate or
other measure of frequency or
association because we don‟t usually
know the „true value‟
 Look for sources of bias in the study
methods used
Selection Bias
 Error due to the way the study participants
are selected
 Occurs when:
   Subjects are not representative of the target
population
   If subjects select themselves
   If subjects in the study are different from those
not in the study
   If the disease/factor under investigation makes
people unavailable for study
Example
 Study looking at drinking habits in
young people.
 Researcher  chose to interview a random
sample of all people 18-25 years (students
Campus, University of Sydney
 Problems with the study??
Sources of Selection Bias
 Poor selection of subjects from the study
population
   non-random selection
    eg volunteers may be different from
those who do not enrol; (important in
cross-sectional surveys; controls in
case control studies)
   ill defined populations eg hospital controls used
in case control study (different criteria for
selecting cases and controls leading to
exclusion bias) e.g. association of reserpine and
breast cancer
Sources of Selection Bias
 Poor selection of subjects from the study
population
   failure to locate or unwillingness to participate
(response bias – those who agree to
participate may be in some way different from
those who refuse to participate or are not
contactable)
   loss due to health outcome (healthy worker
effect, survival bias)
Sources of Selection Bias
Non-participation can be due to:
 Lack of interest or personal relevance

 Inconvenience

 Avoidance of discomfort

 Financial cost

 Antipathy to research
Sources of Selection Bias
Healthy worker effect/bias:
 Usually associated with environmental /
occupational studies
 Low rates of morbidity and mortality found in
both exposed and unexposed in the workplace
 To be active at work – need to be reasonably
healthy
  excludes people who are unhealthy.
Sources of Selection Bias
 Non-random assignment of exposure -
where a doctor allocates treatment or
patients select which group they belong
to
 Omission of subjects from analysis
 loss to follow-up – those who are lost to
follow up or withdraw from the study may
be different from those who are not
followed for the entire study
 inability to obtain adequate measurements
eg missing data
Information Bias

 Systematic error in the measurement
outcome
 Error due to the incorrect
classification of exposure or disease
status
Types of Information Bias:
 Recall bias (or subject error)– those with
a particular outcome or exposure may
remember events more clearly or amplify
their recollections (case-control studies)

Eg mothers of children with birth defects

 Instrument error – problem with the
instrument eg ear temperature instrument
– not calibrated
Types of Information Bias ctd:
 Follow-up or surveillance bias (Cohort
studies) – group with known exposure or
outcome may be followed more closely or
longer than the comparison group
 Hawthorne effect - people act different if they
know they are being watched
 Observer bias (Rosenthal effect) – observers
may have preconceived expectations of what
they should find in an examination
 Interviewer bias - an interviewer‟s
knowledge may influence the structure of
questions and the manner of presentation,
which may influence response
 Intra-observer versus inter-observer bias
Types of Information Bias ctd:
 Misclassification error – errors are made
in classifying disease or exposure status
 Non-differential (random) misclassification
error
 subjects are misclassified with respect to
exposure or case or control in a random
way.
 Subjects have an equal chance of being
misclassified
 This weakens the association observed
between outcome & exposure, if a real
association exists
Example
 Assume that the true incidence rates of
measles in
   non-vaccinated children is 10/1000 person
years
   1/1000 person years in vaccinated children.
   Therefore the theoretical incidence rate ratio
will be 10.
 If 10% of the non-vaccinated children are
misclassified as vaccinated and 10% of
the vaccinated children are misclassified
as not being vaccinated, then the
observed relative rate will be 4.8
 Differential (non random) misclassification
error
 subjects are misclassified with respect to
exposure or case or control in a
non random way

 This changes the direction of the association
observed between outcome & exposure
 E.g. association between alcohol intake and
memory loss. Cases have a greater chance
of being misclassified
“Intention to Treat” analysis in
RCTs

 Analyse groups as they were
originally randomised
“Once randomised always analysed”
 NOT by comparing the control group
with only those in the drug group
who actually took the drug
Choice of Analysis with RCTs
1. Compare those who actually took the
drug and those who actually took the
placebo
This avoids misclassification error
   Potential selection bias
   this ignores the original randomisation
groups
   experimental data  observational data
Choice of Analysis with RCTs
2. Compare those in the drug group to
those in the placebo group (ie the
original randomised groups)
= “intention to treat” analysis
    of what happens in real life.
Our Aim in research is to:

 Reduce bias
 Identify those biases that can‟t be
avoided
 Assess their potential impact
 Use this information when
interpreting results
Controlling Bias

 Choice of the study population and
Control Group

 Select
subjects and controls with
respect to area of residence,
occupation, place of employment
Controlling Bias
Methods of data collection
 Construct specific data collection
instruments eg q‟aires, interviews,
physical examination.
 Use objective closed ended questions
 No differences in implementation
between groups
standardised instruments according to
explicit instructions
Controlling Bias
 Study personnel have standardised
training
 Follow specific procedures identical for all
subjects
 Adopting uniform ways of probing and
responding to questions
 Using standard techniques to deal with
missing information
 Study personnel blinded if possible

 Use multiple sources of data where
possible
Controlling Bias
Selection                      Measurement
Bias                            Bias

Population    Participation Study Follow      Outcome        Analysis
Selection      / Group          up          Assessment
Allocation
High                       Staff trained
participation   No drop outs    in standard    Intention
Representative         rate                       procedures      to treat
No differences                                    Measuring      analysis
between groups    Randomised                     instruments
group                        calibrated
allocation                      No recall
differences

```
To top