There are four reasons besides causality for an association:
1)chance aka random variation
2 Bias aka systematic variation
3)confounding
4)Fraud (typically not even mentioned in statistics books)
One of the 3 most important insights I have read in anything related to medicine came from comments made by Dr. John C Bailard iii in the book he coauthored, "Medical Uses of Statistics"The insight was the recognition of the fragility of medical research in part due to the fact that statistical measures of uncertainty ( i.e p values and confidence limits) particularly in observational studies reflect only part of the " uncertainty that can be attributed to random variability" Medical studies- both randomized trials and observational data- involve "drawing inference from information that inevitably subject to error " that is not controlled with p values and confidence intervals
In one chapter he discussed two articles that appeared back to back in the New England Journal of Medicine in 1983.They were two observational studies dealing with the issue of the cardiovascular effects of hormone replacement therapy in post menopausal women. One study found that the incidence of a type of heart disease was twice as common in the hormone treated group versus the control. The other study found that the hormone treated group have half of the incidence of the control group.
Bailar wrote an editorial in the same issue in which he explained that he had carefully reviewed both papers and could find no flaws in the design or analysis of the data.He admitted he did know why the results were so different.
Life is short, the art long ,experience fallacious and judgement difficult.
No comments:
Post a Comment