News:ASA discusses limitations of p-values what are the alternatives?

Question

[dariober]

Hi All,

I wanted to point out this thread on Cross Validated and the links therein discussing the interpretation of p-values.

http://stats.stackexchange.com/questions/200500/asa-discusses-limitations-of-p-values-what-are-the-alternatives

For those unfamiliar: Cross Validated is the equivalent of Biostars/StackOverflow for statistics. ASA is American Statistical Association and the paper in question is The ASA's statement on p-values: context, process, and purpose

I think the topic, which is not new at all, is quite relevant to many of us!

Answer 1

[John]

There is (and probably always will be) two key questions that non-statisticians want answers too. The first is:

"what statistical approach is best?"

which in our domain would be very similar to the question Which Are The Best Programming Languages For A Bioinformatician? -- and we all know the answer to that. Sure, given certain conditions, one might have some pros on paper over another, but truthfully the best language is the one you are most comfortable solving your problems in - which comes from using it a lot. The same is true for stats, but often people don't like this answer, and suggest it would be "solved" if everyone just used his/her favourite method from day 1. I particularly liked this quote from the linked stats.SE thread:

Which brings me to popular question two:

"why doesn't everyone just learn Bayesian statistics?"

Again, translating that to our domain, this sounds a lot like Why are we still using Bam files? And not Cram, HDF5 or improved Bam files? -- which is frustration borne out of having the theoretical ability to perform a practical task better, but not doing so due to legacy reasons and/or inertia. Personally, I think this is a much more interesting question because it has a solution. We COULD teach everyone on the planet probabilistic stats, but with the current teaching tools available it would take more time than would be worth it. This is why i particularly like this style of learning stats, since it is particularly helpful for bioinformaticians.

[Istvan Albert]

I will answer the first question IMO we are not using CRAM because it does not improve enough on BAM files to make it worth putting up with another format.

Same with bayesian statistics - I think it is an illusion that it is easy to use. I've never seen a bayesian logic explained for a realistic problem. It is always about coin tosses and other overly simplistic examples. I want to see an easy to explain bayesian stat for trusting differentially expressed genes.

I like the idea that we need to learn to embrace uncertainty. And that of course runs through reproducibility. I define that not in the typical manner of using the tool yet thing again but as how easy is to reproduce the finding of a paper with a different method.