Interpreting results of observational researchBMJ 1994; 309 doi: https://doi.org/10.1136/bmj.309.6962.1158 (Published 29 October 1994) Cite this as: BMJ 1994;309:1158
Tools may be blamed for shortcomings of workers
- J Slattery
EDITOR, - Most statisticians will understand Paul Brennan and Peter Croft's concern over the use of P values and confidence intervals in observational studies.1 The authors' remarks are aimed, however, specifically at studies in which researchers are interested in “a presumed cause or treatment.” This is by no means all studies. Many case-control and cohort studies include variables likely to be associated with several potential causes not explicitly stated, the intention being to exploit confounding to narrow down the range of factors that must be examined. A significant P value or range of odds ratios not including unity is then a useful indication that factors associated with the variable are worth investigating. Common examples of such variables are place of residence, social class, and age.
Even in clinical trials we are often not interested in causality in quite the sense that the authors seem to be using it. For instance, to modify one of their examples, a clinical trial in which the treatment consisted of limiting access to oral contraceptives might find …