L21 Association vs Causation Flashcards
In an observational study, how does one determine if an observed association is a valid one?
Need to rule out three alternative explanations for observed association first:
1) By chance
2) By bias
3) By confounding
Why is chance always a possible explanation for an observed association?
Due to random sampling variability, possibly due to:
1) Inadequate sample size: smaller sample size leads to larger chance variability (i.e. reduced statistical power)
2) use of 95% confidence interval, where it may be too wide, resulting in less precise point estimate OR use of too large a significance level tested against (relating to p-value)
- p-value = probability of observed events occurring by chance alone
Why is bias always a possible explanation for an observed association?
Due to systematic error in study design, conduct or analysis of a study, resulting in a mistaken estimate of an exposure’s effect on the risk of an outcome.
- Once bias is introduced, it CANNOT be fixed; thus it is of critical concern the methodology is sound!
- Two major types: Selection bias & Information / observation bias
Identify & differentiate between selection bias &
information bias.
Selection bias:
- Stems from an absence of comparability between groups studied
- If the way in which exposed & unexposed individuals (for cohort studies), or cases & controls (for case-control studies) were selected, is such that an apparent association is observed – even in reality, exposure & outcome are not associated – the apparent association is the result of selection bias (i.e. false positive).
- e.g. non-response bias
Information / observation bias:
- Stems from incorrect determination of exposure or outcome, or both
- When information is incorrect, there is misclassification
(a) Misclassification of exposure status: reporting bias (e.g. pt reluctant to report past exposure), recall bias (e.g. cases recall past exposure better / in more details than controls), incomplete or inaccurate data
(b) Misclassification of outcome status: limited sensitivity & specificity of diagnostic test, incomplete or inaccurate data - Further classified into differential and non-differential misclassification
Differentiate between differential and non-differential misclassification.
Differential: Rate of misclassification differs in different study groups
- Lead to systematic bias in either direction
Non-differential: Rate of misclassification is similar in different study groups
- Lead to systematic bias towards null value (i.e. H0) by diluting the association towards value of 1
Describe what a confounder is, in the context of observational studies.
Confounding occurs when the estimated measure of association between exposure & outcome is distorted (inaccurate) because of the influence of a third variable that is associated with the exposure and influence the outcome.
- ALWAYS an issue in observational studies
- In RCT, randomisation helps to control for confounding due to similar baseline characteristics across study groups.
A variable is considered a confounder if [ALL three conditions MUST be fulfilled]:
1) It is associated with exposure
2) It is a risk factor for outcome
3) It is NOT an intermediary step in the causal pathway from exposure of interest to outcome of interest.
What are possible ways to control for confounding in the study design & data analysis of an observational study?
1) Study Design:
- Restriction (e.g. if smoking is a potential confounder, restrict subjects to ONLY non-smokers)
- Matching (e.g. in a case-control study in which smoking is a potential confounder, cases & controls can be matched by smoking status, where for each case that smokes, four controls who each smokes are selected)
- Optimal ratio to control confounding is 1 case : 4 controls for sufficient statistical power; beyond 1 case : 6 controls -> may not be worthwhile
2) Data Analysis
- Stratified analysis (e.g. analyse smokers & non-smokers separately to examine association independent of smoking)
- Multivariable logistic regression analysis: using statistical models that examine the potential effect of exposure on outcome while simultaneously controlling the effect of many other variables, by calculating crude RR or OR before determining adjusted RR or OR for interpretation
Bias can be adjusted for in data analysis. True or false?
FALSE!
Only confounding can be adjusted for in data analysis, provided that information on potential confounders are available!
- However, residual confounding CANNOT be ruled our, since you cannot cover for ALL variables under the sun that may be confounders.
What are the six common confounders that should be accounted for in observational studies?
SCRAPS: Sex Co-morbidities Race Age Pathology Socioeconomic factors
How does one determine if the observed association causal in nature, provided the observed association is valid?
Use Bradford Hill criteria:
- Framework of 9 criteria for causal inference
- Hill emphasised that meeting all nine criteria is NOT necessary
ACCESS PTB:
1) Analogy: Association established for similar exposure?
2) Consistency: Reproducible in different studies / populations?
3) Coherence: Coherent with generally known facts of natural history & biology of outcome?
4) Experiment: Presence of RCT? (i.e. difficult for observational studies)
5) Strength of association: Magnitude of OR/RR?
6) Specificity of association: Exposure is associated with only one outcome? (weakest)
7) Plausibility: Coherence with current body of biologic knowledge (i.e. physiology/pathophysiology)?
8) Temporality: Exposure -> outcome? (strongest)
9) Biological gradient: Dose-response relationship present?