F or no F: a summary (Alain de Cheveigne )

Subject: F or no F: a summary
From:    Alain de Cheveigne  <alain(at)LINGUIST.JUSSIEU.FR>
Date:    Wed, 19 Jan 1994 16:21:01 +0100

Thanks to all those that responded! I've condensed the responses into a summary, hopefully without perverting their meaning too much. My query was whether the conventional F-value, degrees of freedom and p-value triplet could acceptably be summarized by the p-value alone. My concern was to cut down detail to make things easier to read. The consensus is "leave them in!". The main reasons are: a) Useful redundancy. The degrees of freedom (more than the F-value) give information about the experimental design that might not be clear otherwise. To make things that clear in the text would require so much explanation as to defeat the original purpose of improving readability. b) They help check that the author knows what he's doing. One journal editor said that half of papers submitted contain errors, such as improper ANOVA models, that can be caught by considering the dfs. c) The F-value and the degrees of freedom allow an indirect estimate to be made of the size of the effect relative to the population variance. d) Convention. Everyone reports them, and expects them to be there. Upsetting the reader by breaking the rules also defeats the purpose of readability. Some people added that they'd like to see still more information: means and standard deviation for each condition, and even some indication of the shape of the distribution, such as histograms, so the reader can be sure that the assumptions behind the analysis are valid. Others objected to the practice of reporting post-hoc p-values. A given level of significance should be chosen from the start, and adhered to, and results should be reported in YES-NO format. Meaning, I suppose, that only the F-value and degrees of freedom are reported for each effect. Someone objected to the practice of reporting F and p values with lots of digits: "instead of saying F = 51.4 surely F = 50 would suffice". Finally a few agreed that if all necessary information about the design is well described elsewhere, then even the p-values can be omitted (assuming the results meet the proper level of significance). Many of the answers were very clear and carefully written. It seems a shame to not reproduce them. If the authors agree, I'll make a compilation and post it in another message. Thanks again! Alain

This message came from the mail archive
maintained by:
DAn Ellis <dpwe@ee.columbia.edu>
Electrical Engineering Dept., Columbia University