I have been wondering about this for a bit and forgive my ignorance. Consider we loath the NHST approach, but value the information the p-value gives. I consider the we have a "perfect" experiment. I apply any statistical test, and have n predictors/variables (no idea why a "perfect" study would have n predictor variables, humor me). I also have the effect-size and intervals. Based on a small pilot study and literature I know what a "small" or "large" effect-size is, what I deem reasonable and what kind of variability I can expect. If I consider the p-value as a merit on its own giving me useful information on the variability and consistency in my data. When correcting for multiple comparisons you lose the information the p-value gives. Considering I do not suggest p < .05 as "significant" or any other stark boundary. What rationale is there to correct the p-value for multiple comparison given I do not apply the NHST approach, but describe the observations in the data related to my question(s)?
Thank you in advance
This suggested reference may be helpful to your research. D.Booth
David Eugene Booth Yes I have the book, did not come across it yet or missed it in the +1000 other books. You have a reference to a page?
Regards
Sect 2.9.7 I believe search for bonferroni in the text. You should find the rest of the references near there. There's a separate section about these methods in SPSS toward the end of the book..I have the book in my phone so it's a little bit hard to search. Best wishes, David Booth
Yes, modern technology, pain in the .... . I read it before, but, it assumes a stark boundary whereby I would assume .05 as there is or isn't. The example uses an alpha level of .053=0.86=1-0.86=0.14 or 14% as measure of FP. However, I am not interested in my FP (forgive my bluntness). I am interested in the information of the variability, not in how often I am wrong if I would suggest I am right. I assume I am always wrong (contrary to Colquhoun, (2014), which suggests that nobody want to make a fool of himself ...% of the time) and describe what I observe in the data. Somehow I distanciate myself from a dichotomous statement, but describe the results in relation to my question.
What if for some variable I personally find it reasonable p < .05.-01 and another p
In the old days, instead of adjusting the p values, we used to adjust the alpha values (e.g. the 0.05 cutoff). Thinking about it that way, if you're not using a cutoff, there's nothing to adjust...
Why would you need p-values at all? Why not simply reporting the observed signal-to-noise ratios?
- Badges
- Science topic
- Similar topics
- Statistics
- Survey