Professional Documents
Culture Documents
Module 5: Bias
Selection Bias
2
Other Common Selection Biases
Volunteer Bias = volunteers for studies generally healthier and more
interested in health
Non-Response Bias = People invited to participate and ignore invitation
are different from those who respond
Exclusion Bias = exclusion criteria creates group of participants not
representative of population that produces cases
Healthy Worker Effect = those who attend work regularly generally
healthier that those who dont, affects occupational epidemiology
studies
Non-Differential
Measurement of disease not different for exposed and unexposed
subjects
Heart disease tested in same way for those with high or low BMI
(look at troponin levels)
Measurement of exposure is not different diseased and non-diseased
subjects
Those with heart disease have height and weight measured in same
way as those without heart disease
Non-differential misclassification = subjects may be equally
misclassified on exposure regardless of their status as a case or
control, or they may be equally misclassified on disease status
regardless of whether they are exposed or not
Differential
Measurements of disease different for exposed and unexposed
Those who are overweight might not fit into MRI machine, disease
must be measured some other way
Measurement of exposure different for diseased and non-diseased
Those with heart disease more likely to be called into hospital, so
height and weight measured at hospital, vs. healthy patients who
self report height and weight
Differential Misclassification = knowledge of subject exposure status
influences how subjects will be classified on disease status (or vice
versa)
3
Other Common Information Biases
Reporting bias = reluctance to report an exposure for social or
psychological reasons
Social desirability bias: under report street drug use
Wish bias: deny exposures to minimize self-blame
Recall Bias
Interviewer Bias
Loss to Follow-Up
Solutions
Standard follow-up procedures for all participants
Minimize loss to follow-up (better accommodate participants so they
dont drop out)
Biases in RCTs
Selection Bias
Randomization
Sequence generation
Allocation concealment
Blinding participants,
personnel, outcome assessors
Incomplete outcome data
Losses to follow up
Blinding participants,
personnel, outcome assessors
Confounding
Types of Errors
6
Smoking is a risk factor for lung cancer, independent of yellow
fingernails
Smoking is not a consequence of yellow fingernails, not on causal
pathway
Smoking is a confounder
Yellow fingernails not a confounder, as it is a consequence of
smoking, and yellow fingernails independent of smoking does not
cause lung cancer
Handling Confounding
Design Stage
Randomization balances known
and unknown confounders
(obviously cant be done for
observational study)
Restriction: limit subjects to
those who only fall in one
category of confounder (e.g
include young cows only)
Reduces number of study
subjects
Residual confounding if
category not narrow enough
(young cows vs. very young
cows)
Cannot study influence of
other categories of the
confounder
Matching: only pick controls
with comparable characteristics
to cases
Time consuming and costly
Cannot evaluate effects of
matched variables
Analysis Stage
Multivariable regression
analysis = add confounding
variables into simple regression
models
Must ensure youve
measured the confounder in
the sample
Stratification (same shit as
above)
Computationally difficult
with multiple confounders
(many strata thus many
tables)
Impossible when confounder
is a continuous variable that
cannot be categorized
Identifying Confounders
Effects of Confounding
If
measure of
association differs for each category, then the third variable is an
effect modifier
Mantel-Haenszel Estimator