Upload
vanque
View
214
Download
0
Embed Size (px)
Citation preview
02/11/2014
1
Methods in epidemiology – Medical statistics – Bias
Systematic errors (biases)
Methods in Epidemiology
Methods in epidemiology – Medical statistics – Bias
At the end of the lecture students should be able
� to detail the main types of bias and how they can affect
the interpretation of results
Medical statistics
WWW.SUNHOPE.IT
02/11/2014
2
Methods in epidemiology – Medical statistics – Bias
Actual study
Implementation
Data
Study sample
Planning
Target population
Study population
Study question
Study protocol
Study object Variables to be measured
Structure of clinical research
Inference
Parameter Estimate
iθθθθ̂
Study conclusions
Tθθθθ ( )
iSEθθθθθθθθ ˆ=
Methods in epidemiology – Medical statistics – Bias
Accuracy of conclusions
( )si
E θθθθθθθθ =ˆ iθ̂
ES
� Any conclusions on is
based on the sample estimatei
θθθθ̂
� Actually the ‘true’ effect we are
looking for is Tθθθθ
� We use information from sampling
distribution to infer conclusions on the
‘true’ effect based on the single estimates
actually observed
( )TSi
E θθθθθθθθθθθθ ==ˆ
� Our conclusions will be accurate if
( )si
E θθθθθθθθ =ˆ
iθ̂
ES
( )TSi
E θθθθθθθθθθθθ ==ˆ
WWW.SUNHOPE.IT
02/11/2014
3
Methods in epidemiology – Medical statistics – Bias
( )Si
E θθθθθθθθ =ˆiθ̂
SE
� is the assumption
for a correct inference
� Conversely, if ,
the assumption is unmet and the
conclusion is inaccurate
Tθθθθ
Random error
Systematic error
Accuracy of conclusions
( )si
E θθθθθθθθ =ˆ iθ̂
ES
( )TSi
E θθθθθθθθθθθθ ==ˆ
( )TSi
E θθθθθθθθθθθθ ≠=ˆ
� when both
systematic and random errors
are present
( )TSi
E θθθθθθθθθθθθ ≠=ˆ
Methods in epidemiology – Medical statistics – Bias
True effect +
Systematic error +
Random error =
Observed effect
True effect +
Random error =
Observed effect
( )Si
E θθθθθθθθ =ˆiθ̂
SE
Tθθθθ
Random error
( )si
E θθθθθθθθ =ˆ iθ̂
ES
Systematic error
Accuracy of conclusions
WWW.SUNHOPE.IT
02/11/2014
4
Methods in epidemiology – Medical statistics – Bias
Systematic errors (bias)
Bias results in false understanding about true differences between groups and generates misleading patterns of disease.
May occur in all steps: planning and implementation of studies, data analysis and interpretation of results
Methods in epidemiology – Medical statistics – Bias
Actual study
Implementation
Data
Study sample
Planning
Target population
Study population
Study question
Study protocol
Study object Variables to be measured
Structure of clinical research
Inference
Parameter Estimate
iθθθθ̂
Study conclusions
Tθθθθ ( )
iSEθθθθθθθθ ˆ=
Systematic errors
Randomand
Systematic errors
WWW.SUNHOPE.IT
02/11/2014
5
Methods in epidemiology – Medical statistics – Bias
� Selection bias
� Information bias
� Confounding
Are systematic errors more
likely in observational or
experimental studies?
Bias results in false understanding about true differences between groups and generates misleading patterns of disease.
May occur in all steps: planning and implementationof studies, data analysis and interpretation of results
Systematic errors (bias)
Methods in epidemiology – Medical statistics – Bias
Selection bias
Methods in Epidemiology
WWW.SUNHOPE.IT
02/11/2014
6
Methods in epidemiology – Medical statistics – Bias
Selection bias
� Patients’ selection
Sampling is based on convenience rather than on representativeness, e.g. volunteers, specific sources of subjects, captive populations, health records.
A systematic errors that stems from the procedures used to select patients and from factors that influence study participation.
The association between exposure and outcome differs between participating and not participating subjects
Methods in epidemiology – Medical statistics – Bias
Hutchins et al. 1999
US (≥ 65)
Hutchins et al. 1999
SWOG (≥ 65)
15
Langer et al. 2002
ECOG (≥70)
Selection bias?
Perrone et al. 2002
(=>70)
53
Eld
erl
yp
azie
nts
Elderly patients in clinical trials
WWW.SUNHOPE.IT
02/11/2014
7
Methods in epidemiology – Medical statistics – Bias
Selection bias
� Patients’ selection
Sampling is based on convenience rather than on representativeness, e.g. volunteers, specific sources of subjects, captive populations, health records.
Is related to people. Stems from the procedures used to select patients and from factors that influence study participation.
The association between exposure and outcome differs between participating and not participating subjects
� Study participation
� Compliance with participation (patient’s preferences, drop-out)
� Different chances of being admitted to hospital (Berkson’s bias)
� Different chances of being included in the analysis
Methods in epidemiology – Medical statistics – Bias
Berkson’s bias
General population Subjects admitted to hospital
in the last 6 months
Respir
Dis
Musculo skeletal
dis
YES NO Total
YES 17 207 224
NO 184 2376 2560
Total 201 2583 2784
Association between respiratory diseases and musculoskeletal disorders
Respir
Dis
Musculo skeletal
dis
YES NO Total
YES 5 15 20
NO 18 219 237
Total 23 234 257
8.5% 8.0% 8.0% 21.7% 6.4% 7.8%
29.4% 7.2%
9.8% 9.2%
P = 0,009
WWW.SUNHOPE.IT
02/11/2014
8
Methods in epidemiology – Medical statistics – Bias
585 (26,8%) children did not undergo the tubercolin test:
107 because were not at school the day of test and 478
because their parents did not give consent
Maselli et al. (Monaldi Arch Chest Dis 1997) studied the
prevalence of tuberculosis infection in a random sample of
2.182 children in the city of Napoli.
Missing data:
� not informative (do not affect conclusions)
� informative (may affect conclusions)
Methods in epidemiology – Medical statistics – Bias
Information bias
Methods in Epidemiology
WWW.SUNHOPE.IT
02/11/2014
9
Methods in epidemiology – Medical statistics – Bias
Information bias
� Inaccurate measurements devices
� Inequalities in healthcare or in efforts or time to collect data between the compared groups (biased follow up)
� Discrepancy in obtaining exposure information after disease has occurred (recall bias)
Is related to variables. A subject is misclassified when is placed in an incorrect category of exposure or outcome. It may be:
• differential when misclassification of exposure (outcome) is different for those with and without disease (exposure)
• nondifferential when misclassification of exposure (outcome) is unrelated to the presence of disease (exposure)
Methods in epidemiology – Medical statistics – Bias
WWW.SUNHOPE.IT
02/11/2014
10
Methods in epidemiology – Medical statistics – Bias
Information bias
� Inaccurate measurements devices
� Inequalities in healthcare or in efforts or time to collect data between the compared groups (biased follow up)
� Discrepancy in obtaining exposure information after disease hasoccurred (recall bias)
Is related to variables. A subject is misclassified when is placed in an incorrect category of exposure or outcome.
� Interpretation and reporting of results guided by the researcher’s interests (interpretation bias)
Methods in epidemiology – Medical statistics – Bias
WWW.SUNHOPE.IT
02/11/2014
11
Methods in epidemiology – Medical statistics – Bias
Information bias
� Inaccurate measurements devices
� Inequalities in healthcare or in efforts or time to collect data between the compared groups (biased follow up)
� Discrepancy in obtaining exposure information after disease hasoccurred (recall bias)
Is related to variables. A subject is misclassified when is placed in an incorrect category of exposure or outcome.
� Interpretation and reporting of results guided by the researcher’s interests (interpretation bias)
� Selective reporting of results (publication bias)
Metodologia clinica 5.4 Le persone
WWW.SUNHOPE.IT
02/11/2014
12
Methods in epidemiology – Medical statistics – Bias
Methods in epidemiology – Medical statistics – Bias
Confounding
Methods in Epidemiology
WWW.SUNHOPE.IT
02/11/2014
13
Methods in epidemiology – Medical statistics – Bias
Confounding
Is related to group comparison.
Methods in epidemiology – Medical statistics – Bias
The effect of the exposure on a given outcome is
the difference between the observed progression
of the disease as a consequence of the exposure
and the progression that would have been
observed if subjects would not have been exposed
What is an exposure effect?
This is a counterfactual argument and the exposure effect may not be directly assessed
WWW.SUNHOPE.IT
02/11/2014
14
Methods in epidemiology – Medical statistics – Bias
EXPOSURE
(indipendent of researcher)
STUDY POPULATION
(exposed + not exposed)
STUDY POPULATION
(not exposed)
EXPOSURE
(determined by researcher)
Observational and experimental studies
exposed not exposed exposed not exposed
The exposure effect is estimated by the comparison of the observed effects in exposed and not exposed subjects
To ensure that estimates be accurate and effect be attributable toexposure the compared groups must have similar characteristics
Methods in epidemiology – Medical statistics – Bias
Is related to group comparison. The effect of the exposure is
mixed together with the effect of one or more other variables
leading to bias.
Confounding
WWW.SUNHOPE.IT
02/11/2014
15
Methods in epidemiology – Medical statistics – Bias
Exposure DiseaseThe exposure is
associated with disease
Confounding variable
Confounding
Is related to group comparison. The effect of the exposure is
mixed together with the effect of one or more other variables
leading to bias.
Methods in epidemiology – Medical statistics – Bias
Smoking
(true determinant)
Alcohol
assumptionLung cancer
Smoking is a confounder for alcohol assumption
Smoking (the true determinant) is associated with both alcohol
assumption (exposure) and lung cancer (outcome)
WWW.SUNHOPE.IT
02/11/2014
16
Methods in epidemiology – Medical statistics – Bias
Smoking
(true determinant)
Alcohol
assumption
Lung cancer
Alcohol assumption is not a confounder for smoking
Alcohol assumption is associated with smoking (the true
determinant) but not with lung cancer (outcome)
Methods in epidemiology – Medical statistics – Bias
Confounding
� must be inbalanced between the exposure groups to be
compared
� must be associated with the disease
� must be associated with the exposure
� must not be an effect of the exposure
A confounder may be a true determinant or only a proxy or a
marker an unknown true cause
May be prevented by randomization or controlled by
stratification
WWW.SUNHOPE.IT
02/11/2014
17
Methods in epidemiology – Medical statistics – Bias
Healthy-user effect
Women assuming hormone
replacing therapy had a higher
cultural and social level
> propensity to
use HRT< risk of
CHD
Methods in epidemiology – Medical statistics – Bias
Inconsistent results
between observational
and experimental
studies may follow a
not complete
adjustment by social
level in the analysis
WWW.SUNHOPE.IT
02/11/2014
18
Methods in epidemiology – Medical statistics – Bias
Methods in epidemiology – Medical statistics – Bias
Kigozi et al. (BMC Health Serv Res 2011) estimated the prevalence of HIV in
600 TB patients. Only 405 consented to be tested. Baseline patients’
characteristics are reported in the table by HIV testing. Could missing data
affect the HIV prevalence estimate?
WWW.SUNHOPE.IT