bias and confounding
Checking Date & Time
Reading File
Finding Sources
Finding Sources
Finding Sources
Reading File
Finding Sources
Reading File
Reading File
Reading File
Reading File
Reading File
Reading File
Here is a comprehensive overview of bias and confounding drawn from multiple medical textbooks:
Bias and Confounding in Clinical Research
Both bias and confounding are forms of systematic error that can distort study results and lead to invalid conclusions. They differ in their nature and require different strategies to address.
BIAS
"Bias is a systematic error in the design, conduct, or analysis of a study that can result in invalid conclusions." — Berek & Novak's Gynecology
Bias occurs when compared components are not sufficiently similar — Cummings Otolaryngology. It is usually introduced at the study design stage and is often difficult or impossible to correct analytically afterward.
Types of Bias
1. Selection Bias
Occurs when there is a systematic error in how participants are selected into the study. This creates comparison groups that are not truly comparable at the outset.
- Example: Comparing surgical resection vs. chemoradiation where surgeons avoid enrolling patients with kidney/liver failure — the surgical group ends up with systematically sicker patients, distorting survival comparisons.
- Example: A study on preventive induction where the "induction" physicians were trained in family medicine vs. the control physicians being OB/GYN specialists — the patient populations were inherently different.
- Reduction: Randomization, adequate sample size, careful documentation of non-participants.
2. Information Bias
Occurs when there is a systematic error in how exposures or outcomes are measured or classified.
- Observer bias: Data collected differently across comparison groups.
- Recall bias: In retrospective studies (especially case-control), cases who experienced a bad outcome are more likely to recall or report past exposures than controls. Classic example: mothers of children with cleft lip may more readily recall medication use during pregnancy than mothers of unaffected children.
- Socially desirable response bias: Participants answer based on assumptions about investigators or social norms.
- Reduction: Blinding (single or double), prospective data collection, standardized measurement tools.
3. Collider Bias (Collider-Stratification Bias)
A less intuitive but important type. A collider is a variable independently caused by both the exposure and the outcome. If you inappropriately control for a collider (by design or in analysis), it can induce a spurious (false) association between exposure and outcome.
- Classic example — the "obesity paradox": In general populations, obesity increases mortality. Yet in studies restricted to patients with cardiovascular disease, obesity appears protective. This is because both obesity and cardiovascular disease independently cause mortality; conditioning on cardiovascular disease creates a distorted association.
- Prevention: Causal diagrams (DAGs — directed acyclic graphs) help distinguish confounders (which should be controlled) from colliders (which should not be controlled).
4. Stage Migration Bias (Will Rogers Phenomenon)
When more sensitive diagnostic technology detects disease previously undetectable, patients are "migrated" from earlier stages to later stages. This improves stage-specific survival in both groups without any real change in overall survival. — Cummings Otolaryngology
5. Publication Bias
Negative studies (failing to find a difference) are less likely to be submitted or accepted for publication, skewing the available literature toward positive results. — Miller's Anesthesia
CONFOUNDING
"Confounding occurs when a variable thought to cause an outcome is actually not responsible, because of the unseen effects of another variable." — Cummings Otolaryngology
Criteria for a Confounder
A variable is a confounder if it meets all three conditions:
- Associated with the exposure
- Associated with the outcome
- Not an intermediary step in the causal pathway between exposure and outcome
Classic Example
An investigator claims nicotine-stained teeth cause laryngeal cancer. Despite a strong statistical association, this is spurious — cigarette smoking is the true culprit. Smoking causes both stained teeth (the "exposure") and laryngeal cancer (the outcome). Smoking is the confounder.
Types of Confounders (by measurability)
| Type | Description |
|---|---|
| Known and measured | Can be statistically adjusted |
| Known and unmeasured | Cannot be corrected analytically |
| Unknown and unmeasured | Cannot be corrected at all |
— Miller's Anesthesia
Confounding in Practice
- Observational studies are most vulnerable because groups are not randomly assigned.
- Randomized controlled trials (RCTs): Randomization distributes both known and unknown confounders evenly between groups, which is why RCTs are considered the gold standard for causal inference.
- Residual confounding: Even well-designed observational studies may have unmeasured or poorly measured confounders that remain after adjustment.
CONTROLLING FOR CONFOUNDING
| Method | Stage | Addresses |
|---|---|---|
| Randomization | Design | Known + unknown confounders |
| Restriction (limit study population) | Design | Specific known confounders |
| Matching | Design | Specific known confounders |
| Stratification | Analysis | Known measured confounders |
| Multivariable analysis | Analysis | Multiple known confounders simultaneously |
| Propensity scores | Analysis | Known measured confounders (in observational studies) |
"A propensity score is the probability that a study subject would receive the exposure or treatment of interest, based on the subject's characteristics and the clinical environment." — Goldman-Cecil Medicine
Key limitation: All analytical techniques depend on having data about relevant covariates. Unknown or unmeasured confounders cannot be adjusted for statistically.
KEY DISTINCTION: BIAS vs. CONFOUNDING
| Feature | Bias | Confounding |
|---|---|---|
| Mechanism | Systematic error in process (selection, measurement) | A third variable distorting the exposure–outcome relationship |
| Introduced by | Study design, data collection, analysis | Unequal distribution of a related variable |
| Can it be corrected analytically? | Usually no (especially selection bias) | Often yes (stratification, multivariable analysis) |
| Prevented by | Randomization, blinding, prospective design | Randomization, matching, statistical adjustment |
Sources: Cummings Otolaryngology Head & Neck Surgery | Creasy & Resnik's Maternal-Fetal Medicine | Goldman-Cecil Medicine | Miller's Anesthesia | Berek & Novak's Gynecology
This is a shared conversation. Sign in to Orris to start your own chat.