Based on David_S' EpiGram, I tried to make a decision tree for study design selection.
Which one is better between sample 1 and sample 2?
What do you think about it?
Any better idea?
R.E.
Monday, May 5, 2008
Monday, April 28, 2008
Sunday, April 27, 2008
Correlation does not imply causation
Correlation does not imply causation
The search for causation is central to epidemiology, and all sciences. Often causation is not directly evident from available data. Correlation then is crucial to any consideration of causation. However, correlation, while necessary, is not sufficient for establishing causation. Granted, as Tufte pointed out, it often is a good hint. Qualitative, weight-of-evidence considerations as those involving Hill-like criteria (e.g., plausibility, temporality, specificity, coherence with existing theory) must also be applied.
Just distinguishing causation from correlation can be problematic. To say an event (cause) produces a result (effect) is circular containing definitional dependence. ‘Correlation’ and ‘causation’ are defined at different levels of abstraction. Correlation is often established by statistical similarity, as two variables representing events changing together with unlikely-by-chance regularity. While ‘causation’ has logical and philosophical considerations often referenced to a counterfactual.
For randomized experiments and clinical trials, correlation may generally be closer to causation than in observational studies, with less rigorously generated observations. Subject equality as through randomization helps in eliminating meaningless associations. However, caution is always a caveat to claiming causation. Examples of nonsensical associations abound (e.g., the increase in autism is associated with the increase in home schooling.)
The search for causation is central to epidemiology, and all sciences. Often causation is not directly evident from available data. Correlation then is crucial to any consideration of causation. However, correlation, while necessary, is not sufficient for establishing causation. Granted, as Tufte pointed out, it often is a good hint. Qualitative, weight-of-evidence considerations as those involving Hill-like criteria (e.g., plausibility, temporality, specificity, coherence with existing theory) must also be applied.
Just distinguishing causation from correlation can be problematic. To say an event (cause) produces a result (effect) is circular containing definitional dependence. ‘Correlation’ and ‘causation’ are defined at different levels of abstraction. Correlation is often established by statistical similarity, as two variables representing events changing together with unlikely-by-chance regularity. While ‘causation’ has logical and philosophical considerations often referenced to a counterfactual.
For randomized experiments and clinical trials, correlation may generally be closer to causation than in observational studies, with less rigorously generated observations. Subject equality as through randomization helps in eliminating meaningless associations. However, caution is always a caveat to claiming causation. Examples of nonsensical associations abound (e.g., the increase in autism is associated with the increase in home schooling.)
Finding nothing is finding something
Finding nothing is finding something
Study rationale is rarely conceived in a vacuum. Temptation exists to overprescribe meaning when little exists. ‘Look at these data and find nothing’ seems wry as study advice. But doing so may protect against conclusions based on selection bias, misclassifications, omitted variables, and limited power. Often truth is revealed only from multiple studies and re-hypothesizing, with single studies raising more questions than answered. As with any precept, the pendulum can swing too far in the opposite direction. Ignoring strongly suggestive evidence may cause harm. Above all, contribute to understanding.
Study rationale is rarely conceived in a vacuum. Temptation exists to overprescribe meaning when little exists. ‘Look at these data and find nothing’ seems wry as study advice. But doing so may protect against conclusions based on selection bias, misclassifications, omitted variables, and limited power. Often truth is revealed only from multiple studies and re-hypothesizing, with single studies raising more questions than answered. As with any precept, the pendulum can swing too far in the opposite direction. Ignoring strongly suggestive evidence may cause harm. Above all, contribute to understanding.
Subscribe to:
Comments (Atom)