L3: Bias and confounding Flashcards
Internal and external validity?
Internal validity
Selection bias
Self-selection bias
Diagnostic bias
Information bias
Differential Misclassification
Recall bias
Nondifferential Misclassification
Confounding
External validity or generalizability
Internal validity is a pre-requisite for external validity. When talking about external validity always talk about internal validity. When the study minimises selection bias, information bias, confounding variables = good internal validity now need to think about external validity. No such thing as a perfect study - try to minimise.
what is bias?
bias refers to systematic error in the study design (can be in level of selection, how exposure and outcome is measured, confounders etc.) that results in an estimate of the association between exposure and outcome that is different from the causal association.
Sources of bias
Selection of participants
Measurement aspects
(e.g., exposure, outcome,
confounders) on the
selected participants
what is validity and its types?
Validity of the inferences as they pertain to members of the source population
Prerequisite for external validity
External validity / Generalizability
Validity of the inferences as they pertain to people outside that population
🡺 Validity = lack of bias
Only unbiased procedures are valid
internal validity?
Internal validity
Three general types of biases:
Selection bias
Information bias
Confounding
Have to talk about this when talking about external validity, cannot comment on generalisability without commenting on these.
selection bias?
Distortions that result from procedures used to select subjects and from factors that influence study participation
The relation between exposure and disease is DIFFERENT for those who participate and those who should be theoretically eligible for study (including nonparticipants)
Selecting the wrong controls= bias
Particular problem in case-control studies because the selection of cases and controls (which takes place after outcome has occurred) may be related to exposure
Can also be problem in retrospective cohort studies
self-selection bias?
Self-selection and self-referral can cause bias because the people who choose to participate in a study (or seek out services) may differ systematically from those who don’t. This distorts the results, making them less applicable to the larger population.
Self-referral is considered a threat to validity since the reasons for self-referral may be associated with the outcome.
“Healthy-worker effect”
Healthy people more likely to be working whereas those who remain unemployed, retired, or are disabled, are as a group less healthy.
People might be healthier than others in the general population of that age group
diagnostic bias?
Could fall under selection bias or information bias
Can occur if outcome in individuals is more likely to be ascertained as a consequence of a particular exposure
Example: OC use and venous thromboembolism
Physicians were aware of a possible relationship
Proportion of the women in the study had been hospitalized for evaluation of this disease because they were currently taking OCs
Diagnostic bias e.g: doctors already looking for venous thromboembolism so diagnosis would be higher.
Hormone treatment and risk of heart disease: doctors alreadya ware they are looking for heart disease so diagnosis of heart disease would be higher?
information bias?
Once subjects are in study, the information to be compared between groups can lead to bias if there are errors in the measurement of subjects. Information bias occurs when the data collected about exposure, outcome, or other study variables is inaccurate or misclassified, leading to distorted results. This can happen due to errors in measurement, recall, or reporting.
Types of information bias:
Differential misclassification
Nondifferential misclassification
When talking about information bias always talk about the exposure AND the outcome.
differential misclassification?
Differential misclassification occurs when the errors in classifying an exposure or outcome are different between comparison groups. This means the misclassification is not random but is influenced by the exposure or outcome status, leading to biased results. Particularly likely when exposure is reported after outcome.
For example, in a case-control study on smoking and lung cancer:
If people with lung cancer (cases) are more likely to recall and report their smoking history accurately than those without cancer (controls), this creates differential misclassification of exposure.
If doctors diagnose lung cancer more aggressively in smokers than in non-smokers, this creates differential misclassification of outcome.
This type of misclassification can overestimate or underestimate the true association between exposure and disease, making the results unreliable.
Classification error that depends on the values of other variables.
More of a problem for case-control studies because classification of exposure occurs after disease has occurred
- Exposure Misclassification (Recall Bias)
A study asks mothers if they took certain medications during pregnancy.
A mother whose baby was born with a health issue might remember and report taking medication more carefully than a mother with a healthy baby.
This creates differential misclassification of exposure because mothers of sick babies may recall details differently from those with healthy babies.
- Outcome Misclassification
Suppose smoking increases doctor visits.
If doctors diagnose lung disease more often in smokers just because they see them more frequently, smokers will appear to have higher lung disease rates.
This creates differential misclassification of outcome, making smoking look like a stronger risk factor than it actually is.
- Loss to Follow-up
A study follows people over time to see if physical activity affects mortality.
Less active people (who may be sicker) drop out more often.
This results in fewer deaths being recorded among inactive people, which hides the true risk of inactivity.
reverse causation?
Reverse causation happens when we mistakenly think that an exposure (e.g., low physical activity) causes an outcome (e.g., death), when in reality, the outcome (illness) is what caused the exposure to change.
Example: Physical Activity & Mortality
We see that people who exercise less have higher death rates.
It might look like low physical activity increases the risk of death (A → B).
But what if the real reason people are exercising less is because they are already sick?
In that case, illness (which increases mortality) is actually reducing physical activity (B → A).
So instead of “low activity → higher death risk,” it’s actually “illness → low activity + higher death risk.”
Example: Immune Markers & Cancer
A study finds that people with high immune markers are more likely to be diagnosed with cancer.
We might assume that high immune markers caused the cancer (A → B).
But in reality, cancer might have already started developing before diagnosis, causing immune markers to rise (B → A).
So, instead of immune markers → cancer, it might be cancer (already present) → increased immune markers.
How to Avoid Reverse Causation in Studies?
Careful Selection of Study Population: Exclude people who are already sick at the start of the study.
Delaying the Start of Data Collection: For example, in cancer research, only analyze immune markers from two years after enrollment to reduce the chance that undiagnosed cancer was already influencing immune markers.
Statistical Adjustments: Use models that account for pre-existing conditions to avoid bias.
