Critical Analysis #2 Flashcards
What is Bias?
Bias = the systematic introduction of error into a study that produces results that deviate significantly from the truth.
Cannot be completely eliminated. The aim is to design studies to minimise bias as much as possible.
There are four broad categories of bias:
- Selection bias
- Information/Measurement bias
- Confounding bias
- Attrition bias
Selection Bias
Selection bias occurs during selection of the sample, when the study group doesn’t represent the target population because of the way the sample was selected.
Referral Bias
Type of selection bias.
The referral pathway will determine which group the study participants are selected from.
Spectrum Bias
Type of selection bias.
Occurs when only patients with classical or severe symptoms are selected for a study.
Detection Bias
Type of selection bias.
Preferential selection based on exposure to a particular risk factor.
Non-respondent Bias
Type of selection bias.
Selection based on response to surveys.
Membership Bias
Type of selection bias.
Selection based on membership of certain organisations.
Berkson Bias
Type of selection bias.
Controls selected from hospital, and thus may share baseline characteristics that differ from the study participants.
Ecological fallacy
Inferences derived from groups may not apply to individuals in that group.
Allocation Bias
Type of selection bias that occurs in RCTs.
Reduced by a process called allocation concealment (differs from blinding as occurs at first stage of sample selection).
Inadequate or unclear allocation concealment is associated with a 20-30% exaggeration of treatment effect.
Allocation Concealment
A process to reduce allocation bias in RCTs (a form of selection bias).
Steps in allocation concealment:
1. Generation of a randomised allocation sequence
- Adequate methods include computer generated random numbers, tables of random numbers, drawing lots.
- Inadequate methods include assignment based on DOB, case record number, alternate assignment.
2. Concealment and implementation of sequence
Randomisation
Randomisation can be classified as simple or restricted.
Simple randomisation = based on a single sequence, eg, coin tossing, dice throwing, computer generated random numbers.
Restricted randomisation can be
- blocked (used to obtain groups of equal sizes)
- stratified (used to obtain equal distribution of known prognostic risk factor in groups, useful in small trials to avert imbalances)
- minimisation (ensures balance between groups with respect to several variables, treatment allocated to next participant depends on characteristics of patients already enrolled).
Cluster RCTs
Cluster randomised trials (CRTs) involve randomisation of groups (clusters) of individuals to control or intervention conditions.
The CRT design is commonly used to evaluate non-drug interventions, such as policy or service delivery interventions.
Selection Bias in Cohort Studies
Were only people at risk of outcome included?
Was the exposure clearly defined, specific and measurable?
Were the exposed and unexposed groups similar in all aspects except the exposure?
A national cohort is always best, although often not possible.
The control group can be selected an internal group that develops from the cohort (ie, those exposed who do not develop outcome of interest) or an external control group.
Selection Bias in Case Control Studies
Case and controls should be similar in all aspects except for the disease in question. They should be selected independent of the risk of exposure, so the control group has same risk of exposure as the cases, and is representative of the population at risk of outcome. Selection bias occurs when a control group is inappropriately selected, which is not uncommon.
Ways to minimise selection bias:
- Convenience sample - sample controls in the same way as cases
- Matching - matching cases and controls for known confounders
- Population based sample - random selection of cases and controls from a population based register
Controls can be selected from known groups (random dialing of phone number, nested case control study) or unknown groups (neighbourhood controls, hospital controls, snowball sampling, relatives). A nested case control study involves a group nested within a larger cohort study.
Misidentification Bias
A form of selection bias that can occur in case-control studies.
For example; many suicides may not be classified as such, so the case group is not representative.
Berkson’s Bias
Controls selected from hospital setting.
A form of selection bias.
Neyman Bias
Incidence-prevalence bias. Occurs when a risk factor influences prognosis directly.
For example, smoking kills people before they develop dementia so appears to reduce risk.
Selection Bias in Diagnostic Studies
This occurs when the study population is different from the clinical population in whom the test is applied.
Selection Bias in Meta-Analyses
· Publication bias - studies with positive results are more likely to be published. Can use funnel plot to identify this.
· Language bias - studies with positive results are more likely to be published in English language journals.
· Indexing bias - studies with significant results are more likely to be published in a journal indexed in a literature database.
· Inclusion bias - knowledge of the result of studies may result in differential inclusion of studies
· Multiple publication bias/Citation bias - studies with significant results are more likely to be cited and may be published in multiple journals
Information/Measurement Bias
Information bias occurs during data measurement. It is the error due to incorrect measurement of the exposure and/or outcome.
There are two parties involved, so it can be classified as such:
- Patient = performance/subject bias
- Researcher = observer/investigator bias
Recall Bias
A type of information subject bias.
Plays a role in case control and retrospective cohort studies.
Treatment unmasking
A type of information bias, which can apply to subject or investigator.
The patient becomes aware of the treatment he or she is receiving, for example, due a side effect,
OR the investigator becomes aware of treatment patient is receiving.
Hawthorne Effect
A type of information subject bias.
Participants behave differently when they are aware they are being observed.
Bias to middle and extremes
A type of information subject bias.
Individuals are more likely to give answers close to middle and extremes in rating scales
Social desirability bias
A type of information subject bias.
Patients may choose to give more socially acceptable answers.
Diagnostic/exposure suspicion bias
A type of information observer bias.
The researcher is more likely to search for exposures if they know the disease the patient has, or vice versa.
Outcome measurement bias
A type of information observer bias.
May occur if person measuring outcome is aware of the treatment the patient has received .
Implicit review bias
A type of information observer bias.
Occurs in case control studies when collecting data from notes that may be open to interpretation.
Halo effect
A type of observer information bias.
The knowledge of patient characteristics influences the researcher’s impression of patients with respect to other aspects.
Blinding - Classification of RCTs
Used to minimse information bias in RCTs.
Open label = both patients and investigators are aware of treatment being given.
Single blind = either patients or investigators blinded, but not both.
Double blind = both patients and investigators are blinded.
Triple blind =double blind, plus the outcome assessors are also blinded.
Types of Blinding
- Blinding of patients - less likely to have biased responses to interventions, less likely to seek adjunctive interventions.
- Blinding of trial investigators - less likely to transfer personal attitudes to participants, less likely to administer interventions/adjust dose, less likely to differentially encourage or discourage participants to continue trial.
- Blinding of outcome assessors - less likely to affect outcome measurement.
Minimising information bias in Case Control Studies
Clearly define exposure and blind investigators to outcome status
Minimising information bias in Cohort Studies
Clearly define outcome and blind investigators to exposure status.
Work-up / Verification Bias
A form of information bias in diagnostic studies.
The gold standard test should be applied to all patients irrespective of the diagnosis.
Will Rogers Phenomenon
A form of information bias in diagnostic studies.
Occurs due to improvement in diagnostic tests that can cause different staging of illness across different centres with different diagnostic capabilities.
Minimising Information Bias in Meta-Analyses
Studies are selected based on strict inclusion and exclusion criteria, rating the studies according to quality, with blinding of the investigators here (the investigators rating the studies should be independent from those who selected studies).
Studies are weighted according to allocation concealment, randomisation, blinding, and intention to treat analysis. Weighting of trials is a quality assessment that is ideally done by independent researchers.
For network meta-analysis, similar processes are applied.
Confounding Bias
Error introduced due to confounding factors, which cause a blurring of the association of interest.
Can occur at any stage during the study.
What is a Confounder?
An independent risk factor for the outcome of interest, which is also associated with the exposure. It is not part of the causal pathway directly.
Confounders are only an issue when they occur in uneven distribution between the two groups. That is why randomisation is powerful - it ensures equal distribution of confounders between groups.
Strategies to Reduce Confounding
Design phase = matching for known confounders, randomisation, restriction.
Analysis phase = regression analysis, stratification analysis.
