Lecture 7: Bias and Confounding in Epidemiological Studies Flashcards
How do you get from association to causation?
If bias, confounding, and random error have been ruled out as alternate explanations, then we can get causal inference.
Bias in Knowledge Use
Reporting and Publication Bias
Bias in Analysis and Inference
Information Bias
Selection Bias
Confounding
Random Error
Causal Effect
Types of errors in epidemiological inference
Random Error and Systematic Error
Random Error
- When a sample measurement is different from the true population value due to chance
- Random Error cannot be eliminated but we can reduce the severity through large sample size
Major Sources of Random Error
Individual Biological Variation
Sampling Error
Measurement Error
Individual Biological Variation
- there is a variation in the entity being studied
- caused by intra-individual and inter-individual differences
- ex. blood pressure
- address this error through using repeated measures
Sampling Error
- sample size is too small and not representative of the population
- characteristics of the sample are different due to chance.
- addressed by having a large population size
Measurement Error
- Occurs when there is inaccuracy in instrument or person operating the instrument
- error is due to chance because no measurement is perfectly accurate
Addressed by:
- using strict study protocols
- making sure investigators understand method being used
- making sure labs undergo quality control procedures
What to consider for sample size calculations?
- Level of statistical significance
- Acceptable Error
- Magnitude of Effect Under Investigation
- Prevalence of disease in the population
- Relative sizes of groups being compared
Sample size in case-controls
- if disease is rare, select all cases of disease and sample of controls
- need to achieve a 4 to 1 ratio (that is when the maximum power is reached)
Systematic Error
- Error that is not due to chance
- involved with something that the research has done in regard to the design and implementation of the study
Bias
Systematic error that results in an incorrect/invalid measure of the association
Two major sources of systematic error
Selection Bias
Information/Measurement Bias
Selection Bias
- Systematic differences between people included and not included in the study
- cohort studies: error in selection of exposed and unexposed
- volunteer bias and differential loss to follow up are main sources for prospective
- case-controls: error in selection of cases and controls
- inclusion of controls that are representative of the population for case-control
Selection bias may be due to..
Procedures used to select participants (healthy worker effect)
Factors that Influence continued participation in the study (long follow-up period)
Specific characteristics may affect motivation to participate and ability to participate:
- motivation to participate (worried about exposure)
- ability to participate (low SES, poor health)
Avoiding Selection Bias
Selecting into the study:
- in procedures used to select
- sampling strategy should target person representative of broader population
- incentive to participate
Continuing in the study:
- Keep participants engaged in the research
- ask participants to provide secondary contact
- use tracing method
Measurement Bias
- individuals in the study are systematically placed in the wrong exposure or outcome group
- flaw in measuring exposure, outcome, covariant which results in different information between comparison groups
- Results from systematically imperfect definitions of study variables or data collection methods
Major sources of measurement bias
- recall bias
- observer bias
Recall Bias
- subjects in the study have difficulty recalling information
- mainly in case-controls since you are interviewing the cases about their past controls
- can occur in cohort studies if asked about past exposures
Avoiding Recall Bias
- use objective markers of exposure
- verify responses
- use diseased controls in case-control study (case are more likely to remember information)
Proxy Respondents
- are not the subject themselves but someone close to the subject that can respond on their behalf
- are not going to have the most accurate information as they are not that person
Observation bias
- interview might insert their subjective opinions
- occur in case-control studies as knowledge of disease status might influence questions about exposure
- can occur in cohort and experimental if they inquire about exposed and unexposed
Avoiding Observation Bias
- mask interviewers
- conduct quality assurance activities with interviewers
- test reliability of interviewers’ data
Misclassification
- error in the classification
- non-differential: degree of misclassification is same for study groups
- differential: degree of misclassification is different for study groups
Confounding
- there is a third variable that could be an explanation for the association between the variables under study
- variable is a risk factor for the outcome
- variable is association with the exposure
Effect of confounder
Association may be induced, strengthened, weakened, or eliminated. Impact depends on:
- strength of association between confounder and outcome
- strength of association between confounder and exposure
- prevalence of confounder
Controlling confounders in the study design
- randomization
- restriction
- matching
Limitations of randomization
only used in experimental studies
Limitations of restriction
- difficult to find study subjects
- reduces generalizability
Limitations of matching
- can’t match on more than a few confounders
- difficult to find appropriate controls
- once a variable is matched for, it can no longer be evaluated
Controlling confounders in an analysis design
- stratification
- statistical modeling
Limitations for stratification
- can only control for categorical confounders
- sample sizes may be too small
Limitations of statistical modelling
- difficult for non-researchers to interpret
- requires a large sample size
- certain assumptions need to be met
