Bayesian vs. Frequentist for the applied researcher

trinker

ggplot2orBust
#1
There's been a lot of talk about Bayes lately in my life, talkstats, class, articles etc. I feel like it's becoming something researchers can't hide from anymore. My question is this:

We generally teach applied researchers (ie a content person who learns statistical methods) frequentist statistics and it's generally not to bad for the researcher to understand at an applied level. In comparison, would a Bayesian approach be more difficult to teach to this type of person? Will they be left behind? Will this (and maybe that's the game :)) necessitate the hiring/inclusion of a statistician on a project?

This is a more open question to kind of get a pulse one where we're at and where we're heading.
 

spunky

Can't make spagetti
#2
I feel like it's becoming something researchers can't hide from anymore.

and they shouldn't!!! the Bayesian revolution is on!



My question is this:

We generally teach applied researchers (ie a content person who learns statistical methods) frequentist statistics and it's generally not to bad for the researcher to understand at an applied level. In comparison, would a Bayesian approach be more difficult to teach to this type of person?

i don't think it's necessarily more difficult but i think it can get complicated from the perspective that people need to understand classical/frequentist statistics really well before they can even attempt to compare and contrast them to a different viewpoint of probability. i mean, you're in literacy and i'm in education... both disciplines intersect a lot so we get to talk to people who think similarly... and realise that many of these people don't understand even the very basics of statistics. in my opinion teaching bayesian statistics to this kind of person puts on the instructor a little bit of a challenge in terms of how to make things approachable to the students. i believe the puppy book does a great job at it and, up until what i've found, i think it's maybe the only textbook i'd ever consider using if i were to teach a module of bayesian stats to my colleagues



Will they be left behind?

or maybe they already are.... lol ;) it really depends though. as most of the members on this board have commented on, i dont think the classic ANOVA, regression, t-test stuff we've learned to love will go anywhere. it's here to stay. what i do see, however, is that as journal articles start becoming a little bit more demanding on the complexity of their statistical analyses, perhaps more and more people will need to start considering jumping into this bandwagon because of all the problems it solves when compared to classical statistics...


Will this (and maybe that's the game :)) necessitate the hiring/inclusion of a statistician on a project?

i certainly hope so because i need a better paycheck!!! :p

nah, on a more serious note... so Bayesian statistics in our filed is starting a little bit as the new, cool stuff to do and maybe that's why there's such a buzz around it now. we're usually a little bit late among the sciences to incorporate cool, new quantitative techniques because we need to 'SPSS-ify' the software before anyone even considers using them... and there're always a few detractors that tend to force things to stay the way they are (i really recommend reading the monography written by Tatsuoka about how hard it was to bring multivariate techniques into the social sciences just to give you an idea). now, with that being said, the use of estimation methods/algorithms traditionally attributed to Bayesian statistics (Gibbs sampling, Metropolis-Hastings, etc.) can help solve a lot of problems. for example, a lot of Item Response Theory problems treat items as dimensions across which you need to integrate to obtain probabilities. if you have questionnaires with 200 or 500 questions you have a 200 or 500-dimensional integral where the regular software is just not gonna cut it. that's why we have all this extensions to maximum likelihood and the need for humongous sample sizes to obtain item parameters. Bayesian statistics solve this problem quite efficiently plus it provides a much more direct way to make inferences. i think a major issue, however is that Bayesian statistics relies on people to frame their research questions in terms of statistical models, and we dont teach students how to think in terms of models enough. that's probably going to be one of the challenges in creating an SPSS analogue for Bayesian stats because i dont think it'll be very easy for a GUI-only software to handle all the possible cases that we could find in real life...

PLUS there's always Lazar's accurate observation: if you have a manuscript that has 'Bayesian' in its title it increases your chances of publication by a bazillion percent. i just realized that myself with my thesis. :)