Prevention & care
Causation and confounding: assumptions, selection and competing explanations
On this page
A causal question compares what would happen under different exposures or interventions in a specified population. An association instead compares observed groups. These can differ because the groups have different prognoses, because the outcome influences exposure, or because selection and measurement change the comparison.[1]
Define the target and the role of a variable
| Term | Meaning | Why it matters |
|---|---|---|
| Causal contrast | Compares potential outcomes under specified alternatives. | Define the population, exposure versions, outcome and time.[1] |
| Confounding | An exposure–outcome comparison is distorted by common causes. | Comparable groups require more than observing an association.[2] |
| Mediator | A variable lies along a causal pathway. | Adjusting for it can change a total-effect question.[2] |
| Collider | A variable is a common effect on a path. | Conditioning on it can create an association.[3] |
| Selection bias | Selection into an analyzed group can distort the comparison. | Recruitment, retention and analysis restrictions need examination.[5] |
| Instrument | A variable affects an exposure and meets additional causal conditions. | Relevance alone is insufficient for instrumental inference.[6] |
Make the comparison precise before calculating an effect
The What If framework separates observed associations from comparisons of potential outcomes. Identification generally requires consistency, exchangeability and positivity: the exposure contrast must be meaningful, relevant prognostic differences must be addressed, and the compared strategies must be possible within the groups analyzed. These conditions concern the scientific question and data, not merely a regression command. A hypothetical change in a biomarker may be different from a particular medicine that changes it through several pathways.[1]
A diagram makes assumptions visible rather than proving them
Greenland, Pearl and Robins introduced causal diagrams to help determine which paths an adjustment opens or closes. For example, a common cause of exposure and outcome can create a noncausal association. A variable on the causal pathway plays a different role. Arrows encode substantive assumptions, including the absence of proposed effects. The data do not automatically establish the correct graph. Domain knowledge, timing and plausible alternative diagrams should therefore accompany a proposed adjustment set.[2]
A confounder can change during follow-up
Time-varying confounding occurs when factors influencing later exposure and outcome change over time. A special difficulty arises if earlier exposure also changes those factors. Imagine treatment affecting symptoms, while later symptoms influence both the next treatment choice and the eventual outcome. This is feedback, not just a baseline difference.[9]
Evidence & uncertainty
An alcohol genetics analysis illustrates why alternative designs can challenge familiar associations. Its findings concerned a genetic variant, alcohol-related behavior and cardiovascular outcomes in European-ancestry data. Interpreting those results as an intervention effect still requires assumptions; genetic assignment is not a randomized drinking programme.[7]
Selected-source overview · Independent review pendingPublished questions
Why is an observed group comparison not automatically a causal effect?
Selected-source overview · Independent review pending. Read the evidence context.
The groups may have different prognoses under the same exposure. A causal contrast asks what would happen under specified alternatives; consistency, exchangeability and positivity connect that target to the observed data under assumptions.[1]
Evidence behind this answer
- Causal Inference: What If
original causal methods book · 2020
Original author landing page inspected; latest linked August 19, 2026 PDF downloaded directly via urllib after web retrieval failed; selected substantive chapter passages extracted and inspected locally · Chapters 1–3: potential outcomes, association, consistency/exchangeability/positivity and intervention versions
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Does drawing a causal diagram establish that its arrows are correct?
Selected-source overview · Independent review pending. Read the evidence context.
No. The graph represents assumptions based on timing and substantive knowledge. It helps identify adjustment consequences, while competing graphs or unmeasured common causes can lead to different conclusions from the same observed associations.[2]
Evidence behind this answer
- Causal diagrams for epidemiologic research
original graphical causal methods paper · 1999
Selected original published PDF passages inspected through Ohio State course copy; primary author bibliography confirms publication · Paths, common causes, mediation/conditioning and assumptions in diagrams
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Does controlling every measured variable reduce bias?
Selected-source overview · Independent review pending. Read the evidence context.
No. A mediator or collider may make the estimate answer a different question or create bias. Measurement error and reverse direction also matter. The variables should be chosen for the causal target and their assumed roles.[4]
Evidence behind this answer
- Thinking Clearly About Correlations and Causation: Graphical Causal Models for Observational Data
original causal methods primer · 2018
Selected original published 16-page PDF passages inspected through UBC academic copy; publisher indexed metadata inspected · Control-variable roles, reverse direction, surrogate interventions and external validity
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Did the collider example show that a sandwich prevents influenza?
Selected-source overview · Independent review pending. Read the evidence context.
No. It used hypothetical data. Restricting attention to fever selected on a common effect and created a misleading association between otherwise unrelated causes. The example explains a mechanism of bias, not an infection-treatment finding.[3]
Evidence behind this answer
- Illustrating bias due to conditioning on a collider
original hypothetical methods illustration · 2009
Original published four-page PDF inspected through University of Colorado course copy; direct PubMed retrieval blocked · Fever/sandwich/influenza example and conditioning definition
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Can a huge sample eliminate selection bias?
Selected-source overview · Independent review pending. Read the evidence context.
No. Systematic differences in recruitment, retention or analysis restrictions can persist with more observations. Correction requires appropriate measured information and assumptions; increased sample size alone only improves precision around the available comparison.[5]
Evidence behind this answer
- A structural approach to selection bias
original selection bias methods paper · 2004
Original PubMed indexed abstract and selected original author-hosted published PDF indexed weighting/discussion passages inspected; direct PDF retrieval failed; no full-text appraisal · Selection mechanisms, common-effect structure, weighting and assumptions
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Is every genetic association a valid Mendelian-randomisation instrument?
Selected-source overview · Independent review pending. Read the evidence context.
No. Relevance, independence and exclusion restriction must be considered. Pleiotropy, weak instruments and population structure can create problems. Diagnostic checks can challenge assumptions without proving that every required condition holds.[6]
Evidence behind this answer
- Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians
original instrumental inference methods guide · 2018
Original publisher and PMC indexed substantive methods/table passages inspected; direct full-text retrieval blocked · Three assumptions, weak instruments, pleiotropy/population structure and lifetime-exposure interpretation
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Was the alcohol genetics study a randomized drinking trial?
Selected-source overview · Independent review pending. Read the evidence context.
No. It compared variant-related differences in alcohol behavior and cardiovascular outcomes. The large European-ancestry dataset does not remove instrument assumptions or establish the effect of one particular drink or short-term behavior-change programme.[7]
Evidence behind this answer
- Association between alcohol and cardiovascular disease: Mendelian randomisation analysis based on individual participant data
original individual participant mendelian randomisation analysis · 2014
Original PubMed abstract and original investigator repository/indexed published results inspected; no independent full-text appraisal · 56 studies/261,991 European ancestry; ADH1B 17.2% lower weekly intake and coronary odds 0.90; interpretation limits
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Does target trial emulation make an observational database randomized?
Selected-source overview · Independent review pending. Read the evidence context.
No. It makes the intended comparison explicit and can prevent design errors, including mismatched start times. The emulation still relies on exposure, outcome and confounder information and assumptions that observational data may not satisfy.[8]
Evidence behind this answer
- Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available
original target trial methods framework · 2016
Selected original published seven-page PDF passages inspected through EPFL academic copy; publisher date/indexed abstract inspected · Protocol components, baseline alignment, future selection and measured-confounding limits
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Why is it hard to compare treatments that change as people get better or worse?
Selected-source overview · Independent review pending. Read the evidence context.
Health changes may influence the next treatment and the outcome, while earlier treatment may have helped cause those health changes. Methods must address that feedback; a longer list of adjustment variables is not automatically a solution.[9]
Evidence behind this answer
- Marginal Structural Models and Causal Inference in Epidemiology
original causal methods article · 2000
University-hosted original published 11-page PDF retrieved. Selected definitions, feedback conditions, weighting framework, no-unmeasured-confounding and positivity/model limitations read. Hypothetical symptom example is explanatory, not an observed clinical result. · Original published PDF abstract/introduction; time-dependent confounding conditions; assumptions in Section 4 and limitations in Section 11
This answer summarizes selected sources. An independent appraisal of the complete evidence has not been completed.
Temporal order and the estimand affect adjustment choices
Rohrer’s methods paper explains why adding more control variables is not automatically better. A variable can be a confounder for one question and a mediator or collider for another. An association can also reflect reverse direction: emerging illness may change activity, diet or service use before the illness is diagnosed. Measuring an exposure before diagnosis does not necessarily mean it preceded the disease process. Competing explanations should be specified rather than hidden behind the word adjusted.[4]
A published hypothetical example shows how restriction creates bias
Cole and colleagues use an explicitly hypothetical meeting where influenza and a contaminated sandwich independently cause fever. Before restriction, sandwich assignment and influenza are unrelated. Among only people with fever, someone without the sandwich must have influenza in that simplified world, producing an apparent association. The sandwich did not prevent influenza: selecting on their common effect created the pattern. This is a mathematical illustration, not an observed infection outbreak or evidence about actual diagnostic probabilities.[3]
Who remains in the analysis can change the conclusion
Hernán and colleagues explain a common structure underlying several selection problems, including inappropriate controls, differential loss to follow-up and volunteer selection. Restriction or adjustment involving common effects can distort associations. A larger sample does not repair a systematic selection mechanism. Weighting can sometimes address selection if the relevant variables are measured and the modeling assumptions are adequate. It cannot guarantee recovery when important causes of selection or outcome are absent from the data.[5]
Genetic instruments require three core conditions
The Mendelian-randomisation guide distinguishes relevance, independence and exclusion restriction. The variants must relate to the exposure, avoid relevant common causes, and affect the outcome through the exposure contrast being studied. Weak instruments, population structure and horizontal pleiotropy can undermine interpretation. Checks and sensitivity analyses can reveal problems without proving every assumption. A genetically influenced difference across life also need not equal a brief intervention beginning in adulthood. The analysis needs a stated causal target, not just a significant variant–outcome association.[6]
A genetic analysis challenged a simple protective-drinking interpretation
Holmes and colleagues analyzed 56 studies with 261,991 people of European descent. Carriers of an ADH1B variant drank 17.2% less alcohol per week and had lower coronary-heart-disease odds, about 0.90 relative to noncarriers. The investigators interpreted the pattern as challenging the proposed cardiovascular benefit of light-to-moderate drinking. This design differs from comparing self-selected drinkers with abstainers, which may reflect health and behavioral differences. Its instrument assumptions, ancestry, multiple related alcohol behaviors and exposure across life remain relevant. It is not a trial assigning a particular drink, amount or cessation programme.[7]
Observational treatment comparisons can specify a target trial
Hernán and Robins propose writing the trial one would ideally conduct: eligibility, treatment strategies, assignment, follow-up, outcomes, causal contrast and analysis. Observational emulation must align eligibility, strategy assignment and the start of follow-up. Defining treatment groups using later survival or future treatment can introduce avoidable bias. The framework clarifies design and helps expose problems, but it does not create random assignment or erase unmeasured confounding. A very large database can remain inadequate for the target question.[8]
Causal language should reflect the supported comparison
Rohrer emphasizes that causal inference is not a mechanical conversion of correlations. A lab manipulation of perceived social position, for example, can answer a different question from lifelong material circumstances. Strong internal control does not make those exposure versions interchangeable. Observational analysis can contribute to causal understanding when its assumptions are argued carefully. The useful conclusion identifies what was compared, what alternatives remain plausible and what evidence would distinguish them, rather than declaring every association causal or every observational study useless.[4]
Why one adjustment rule may not settle feedback
Under that feedback structure, simply adjusting for the later factor can remove part of the earlier exposure’s effect, while ignoring it can leave later confounding. This does not mean ordinary adjustment always fails. Marginal structural models are methods designed for such settings; some use weighting. They still require relevant confounding information, suitable models and possible exposure choices across the compared histories.[9]
References
- Causal Inference: What If
Original causal methods book
Miguel Hernán and James Robins; original author-hosted book · Content/publication date: 2020
Accessed: 2026-10-14 UTC · Original author landing page inspected; latest linked August 19, 2026 PDF downloaded directly via urllib after web retrieval failed; selected substantive chapter passages extracted and inspected locally
Cited sections: Chapters 1–3: potential outcomes, association, consistency/exchangeability/positivity and intervention versionsSuggested citation retains 2020; inspected PDF cover/revision August 19, 2026 and copyright 2020/2026. Whole book and historical Harvard/author PDFs count as one underlying publication allowance. Selected chapters, not exhaustive book appraisal.
- Causal diagrams for epidemiologic research
Original graphical causal methods paper
Greenland, Pearl and Robins; Epidemiology · Content/publication date: 1999
Accessed: 2026-10-14 UTC · Selected original published PDF passages inspected through Ohio State course copy; primary author bibliography confirms publication
Cited sections: Paths, common causes, mediation/conditioning and assumptions in diagramsDOI 10.1097/00001648-199901000-00008. Mirrors and earlier technical report are not separate allowances or independent replications. Graph is assumed, not inferred conclusively from correlations.
- Illustrating bias due to conditioning on a collider
Original hypothetical methods illustration
Cole and colleagues; International Journal of Epidemiology · Content/publication date: 2009
Accessed: 2026-10-14 UTC · Original published four-page PDF inspected through University of Colorado course copy; direct PubMed retrieval blocked
Cited sections: Fever/sandwich/influenza example and conditioning definitionAdvance Access November 19, 2009 ; 2010 issue 39:417–420. Explicitly hypothetical examples, not empirical trial rates. DOI 10.1093/ije/dyp334; PMC 2846442 mirrors same publication.
- Thinking Clearly About Correlations and Causation: Graphical Causal Models for Observational Data
Original causal methods primer
Julia Rohrer; Advances in Methods and Practices in Psychological Science · Content/publication date: 2018
Accessed: 2026-10-14 UTC · Selected original published 16-page PDF passages inspected through UBC academic copy; publisher indexed metadata inspected
Cited sections: Control-variable roles, reverse direction, surrogate interventions and external validityPublication year retained; accepted November 9, 2017 is not publication. DOI and repository PDFs remain one report. Hypothetical illustrations and method argument do not establish clinical intervention effectiveness.
- A structural approach to selection bias
Original selection bias methods paper
Hernán, Hernández-Díaz and Robins; Epidemiology · Content/publication date: 2004
Accessed: 2026-10-14 UTC · Original PubMed indexed abstract and selected original author-hosted published PDF indexed weighting/discussion passages inspected; direct PDF retrieval failed; no full-text appraisal
Cited sections: Selection mechanisms, common-effect structure, weighting and assumptionsDOI 10.1097/01.ede.0000135174.63482.43. Original indexed discussion supports reweighting under measured-variable and nonzero-probability conditions. Direct PDF retrieval timed out or returned 502; no full PDF access or invented empirical correction result claimed.
- Reading Mendelian randomisation studies: a guide, glossary, and checklist for clinicians
Original instrumental inference methods guide
Davies, Holmes and Davey Smith; The BMJ · Content/publication date: 2018
Accessed: 2026-10-14 UTC · Original publisher and PMC indexed substantive methods/table passages inspected; direct full-text retrieval blocked
Cited sections: Three assumptions, weak instruments, pleiotropy/population structure and lifetime-exposure interpretationDOI 10.1136/bmj.k601; PMID 30002074; PMC 6041728 all one report. Genetic instruments not assumed automatically valid or equivalent to an adult intervention.
- Association between alcohol and cardiovascular disease: Mendelian randomisation analysis based on individual participant data
Original individual participant mendelian randomisation analysis
Holmes and colleagues; The BMJ · Content/publication date: 2014
Accessed: 2026-10-14 UTC · Original PubMed abstract and original investigator repository/indexed published results inspected; no independent full-text appraisal
Cited sections: 56 studies/261,991 European ancestry; ADH1B 17.2% lower weekly intake and coronary odds 0.90; interpretation limitsDOI 10.1136/bmj.g4164; PMC 4091648 and investigator/publisher copies remain one publication. Lifelong genetically related alcohol behaviors do not prescribe a drink amount or prove a short-term intervention effect.
- Using Big Data to Emulate a Target Trial When a Randomized Trial Is Not Available
Original target trial methods framework
Hernán and Robins; American Journal of Epidemiology · Content/publication date: 2016
Accessed: 2026-10-14 UTC · Selected original published seven-page PDF passages inspected through EPFL academic copy; publisher date/indexed abstract inspected
Cited sections: Protocol components, baseline alignment, future selection and measured-confounding limitsOnline March 18, 2016; April 15 issue. DOI 10.1093/aje/kwv254; PMC 4832051 same work. Protocol hormone-therapy example is methodological, not a newly estimated clinical result.
- Marginal Structural Models and Causal Inference in Epidemiology
Original causal methods article
Robins JM, Hernán MA and Brumback B · Content/publication date: 2000
Accessed: 2026-10-20 UTC · University-hosted original published 11-page PDF retrieved. Selected definitions, feedback conditions, weighting framework, no-unmeasured-confounding and positivity/model limitations read. Hypothetical symptom example is explanatory, not an observed clinical result.
Cited sections: Original published PDF abstract/introduction; time-dependent confounding conditions; assumptions in Section 4 and limitations in Section 11Historical methods paper; assumptions remain necessary.
Scope & review
- Population
- Readers assessing causal health claims and researchers’ explicit assumptions
- Exposure or intervention
- Counterfactual contrasts, common causes/pathways, reverse direction, selection, instruments and target trials
- Comparison
- Observed exposure groups versus defined alternative strategies or variant-related contrasts
- Outcomes
- Association versus causal targets; selected genetic coronary outcomes and hypothetical bias illustration
- Timeframe
- Time ordering, lifetime genetic exposure and specified follow-up; no universal causal time horizon
- Jurisdiction
- International methods education; European ancestry 2014 example and explicitly hypothetical illustration; no personal alcohol or treatment prescription
Selected-source educational explanation. Official guidance, original study findings and editorial interpretation are distinguished. No exhaustive search, independent systematic synthesis or qualified medical review has been completed.
Sources checked on their individual access dates. Selected sources were checked; ongoing surveillance and qualified review have separate statuses on the coverage page.
Revision history
- 2026-10-14 UTC: Added a bounded study and causal methods foundation guide; qualified clinical and research-methods review remains pending.
- 2026-10-20 UTC: Added supported foundational, practical and evidence explanations after reading full target bodies and relevant current/pending adjacent coverage. Original sections, questions and sources preserved as exact prefixes; qualified review pending.
What this article still leaves open
- Which exposure versions and population define the causal claim?
- What selection or unmeasured causes could change the estimate?
- Would another plausible causal structure produce the observed association?
Suggest a correction
Submit a content correction to the editorial queue, or download a local note. Leave out personal medical details. Submitted text is stored privately with a browser receipt session and is visible to this browser and the editor. Clearing cookies removes access to your submission history. This is not an account or a medical service.
Your submitted corrections for this article
Explore the topic map
Browse the health domains connected to this article. The map also identifies topics whose articles remain unwritten.