Error
Precision
Validity
Error = discrepancy between the observed result and the true value
- -typically resulted from: random processes like random sampling, OR systematic processes
- -bias = systematic error
Precision = absence of random error Validity = absence of bias (or absence of all error)
Internal vs External Validity
Internal Validity = Whether the study provides an unbiased estimate of what it claims to estimate
External Validity = Whether the results from the study van be generalized to some other population
Direction of Bias
Risk Exposure
Positive bias = observed value is higher than true value
Negative bias = observed value is lower than true value
Bias towards the null = observed value is closer to 1.o than true value
Bias away from the null = observed value is farther from 1.0 than is the true value
Direction of Bias:
Preventive Exposure
Positive bias - observed value is smaller than true value
Negative bias - observed value is higher than the true value
Bias towards the null - observed value closer to 1.0 than true value
Bias away from the null - observed value is farther from 1.0 than true value
Selection Bias
The way in which subjects are selected into the study population or into the analysis distorts the effect estimate.
Cohort study - disease status influences selection of subjects (more exposed cases are detected than unexposed cases)
Case Control study - exposure status influences selection of subjects
Cross Sectional study - either variable in a student influences the selection of subjects
Selection bias less likely to occur in prospective cohort studies because pple are recruited before cases arise
Selection bias most likely to occur when investigator can’t identify the base population
Self Selection Bias
When the exposed group is selected from group of volunteers
Estimated exposure effect could be biased - volunteers might different in ways related to the outcome
Selective /Differential Loss to F/u
Disproportionate loss of selected subjects during the follow up period.
Attrition may be due to other causes or death, lack of subject cooperation, etc.
In practice - do not have outcome info on lost subjects
we can’t know the direction/magnitude of the bias
The amount of bias may differ considerably for any given amount of attrition
Greater attrition = greater the max possible bias could occur
We cannot determine whether bias occurred simply by comparing exposure distribution
Selective Survival Bias
Occur from the disproportionate loss of potential subjects before selection
If exposure status is associated with the loss of eligible subjects, differentially for cases/noncases
Detection Bias
If certain cases of disease under study never get detected
Because certain pple have access to intensive medical attention, increased likelihood of disease detection
Berkson’s Bias
concerns hospital controls:
- if hospital based cases/controls have different exposures than the base population , the OR will be biased
- hospital based controls may be less likely to have exposures of interest than the population they are supposed to represent, the OR will be over-estimated
Solution - use population based controls
Temporal Ambiguity
Certain study designs and selection strategies can lead to bias if occurrence or presence of disease directly or indirectly affect exposure status
observed results may reflect the effect of disease on exposure rather than effect of E on O
Dealing with Selection Bias
In Planning Stage of Study
- Use incidence data, not prevalence data, when possible.
- Case-control studies: select controls from the actual base population from which study cases arose—i.e., use a population-based design.
– may not be possible or feasible to do with certain diseases in certain situations (e.g., studying a rare disease that is difficult to dx in a population with no systematic surveillance system) - Case-control studies, that are not perfectly population biased, use two or more control groups selected in different ways or from different populations
Ex., a control group of hospitalized patients, + a control group of community residents.
– each control group might introduce a different bias = further complicates interpretation. - Apply the same eligibility criteria for selecting all subjects.
- Make sure that all potential subjects (exposed and unexposed) undergo the same diagnostic procedures and intensity of disease surveillance.
Dealing with Selection Bias
In Data Collection Stage of Study
- Minimize nonresponse, nonparticipation, and loss to follow-up; keep a record of all such losses and collect baseline data on them.
- Collect as much information as possible regarding exposure history, including times and reasons for changes in exposure status.
- Make sure that the disease is diagnosed blind to exposure status.
Dealing with Selection Bias
In Data Analysis Stage of Study
- usually too late
1. Compare nonresponders/dropouts with responders/nondropouts with respect to baseline variables.
***Note, however, that such analyses cannot confirm the presence or absence of bias or the direction of the bias, and they cannot be used to estimate the magnitude of the bias.
- Using study results, prior knowledge, and logic, try to deduce the direction of specific biases and, if possible, estimate the approximate magnitude of these diseases.
Information bias
- Results from imperfect definitions of study variables OR flawed data collection procedures
- Consequences: misclassification of E/O status for a significant proportion of participants, = invalid studies.
Reliability + Validity
Reliability = extent to which the measurement obtained with a particular test or instrument are reproducible or repeatable.
Validity = extent to which measurements reflect the true values of the theoretical factors that the observed variable is supposed to measure
What Constitutes a Valid Study?
What is Sensitivity
What is Specificity
A valid/ unbiased study = based on its design, methods, and procedures, will produce overall results that are close to the truth
Sensitivity + specificity = two main components of validity.
Sensitivity: the ability of a test (or a measure) to identify correctly those who have the disease/outcome
Specificity: the ability of a test (or a measure) to identify correctly those who do not have the disease /outcome
Exposure Identification Bias
occurs when there are problems in the collection of exposure data or an imperfect definition of the level of exposure
Less likely in cohort studies bc exposure is ascertained before outcome occurs
May occur in case-control studies
2 subcategories: recall + interviewer bias
Exposure identification Bias :
Recall Bias
Results from inaccurate recall of past exposure.
More likely in case-control + cross-sectional studies than in cohort studies bc exposure info was asked in the past
How to Prevent Recall Bias
Verification of exposure information obtained from participants by review of pharmacy or hospital charts, or other sources.
Use of objective markers of exposure
Use of the cohort study design, including the conduct of case-control studies within the cohort.
Exposure Identification Bias
Interviewer Bias
Observer bias in ascertaining exposure – may occur if disease status in a case-control study is not masked
i,e, clarifying questions probing, skipping rules in procedures
How to Prevent Interviewer Bias
- Conduct reliability and validity sub-studies in samples
2. Masking of interviewers with regard to case-control status.
Outcome Identification Bias
- May result from either differential or nondifferential misclassification of disease status –> due to imperfect definition of the outcome, or to errors in data collection stage
- May occur in both case-control and cohort studies.
Observer Bias and Respondent Bias
Outcome Identification Bias
Observer Bias
How to Prevent?
Cohort study – outcome ascertainment may be affected by knowledge of the exposure status of the study participant,
**particularly when the outcome is “soft”, such as migraine episodes or psychiatric symptoms.
To prevent: mask observers by exposure status, use multiple independent observers.
