Bayesian vs frequentist for social statistics

[Zaphod2020]

Reviewing the research interests of professors across various departments, I've noticed that statisticians who apply their work in the social sciences tend to be Bayesians. Do social scientific problems lend themselves to Bayesian methods or am I just making this up?

[Bayesian1701]

I am totally biased as someone interested in Bayesian statistics applied to the social sciences.  I think P(Bayesian| Social Science applications) is at least 0.5,  and probably much higher (.7ish) but my priors might be biased.  I don't want to get into why Bayesian is popular in general but I like it because the problem I am interested in has a lot of parameters and great prior information that lends itself to work well with Bayesian methods.

[bayessays]

One reason is likely the ease of incorporating different sources of income together (eg surveys). Another reason is the ease of translating between Bayesian probabilities and decisions that incorporate costs, rather than scientific studies where you might want to just test a hypothesis. 

[Stat Assistant Professor]

Some of the reasons I can think of off the top of my head: 1) Bayesian models allow you to incorporate prior information/beliefs about the parameters of interest into your model. If you have little or no information, you can still perform Bayesian inference by choosing appropriately "weakly informative" or "non-informative" priors. 2) In some ways, the Bayesian interpretation of uncertainty quantification is easier and more "natural." For example, when you talk about confidence intervals in the frequentist sense, you're not talking about probabilities, you're talking about "long-term" coverage (i.e. if you construct many intervals, 95% of them will contain the true parameter). Meanwhile, the Bayesian credible interval gives you a 95% probability that the true parameter is contained in said interval. 

I am not sure about specific applications in social science, but it is sometimes a philosophical difference as to whether it is appropriate to describe your parameters of interest probabilistically (i.e. whether the parameters should be treated as random variables, rather than as fixed values). 

Edited March 7, 2018 by Applied Math to Stat

[Zaphod2020]

Thanks for the replies!

I'd be curious to know P(Bayesian| Social Science applications), as @Bayesian1701 put it.

[Zaphod2020]

14 hours ago, Bayesian1701 said:

I am totally biased as someone interested in Bayesian statistics applied to the social sciences.  I think P(Bayesian| Social Science applications) is at least 0.5,  and probably much higher (.7ish) but my priors might be biased.  I don't want to get into why Bayesian is popular in general but I like it because the problem I am interested in has a lot of parameters and great prior information that lends itself to work well with Bayesian methods.

Would you mind talking about the problem you're interested in?

[TakeruK]

14 hours ago, bayessays said:

One reason is likely the ease of incorporating different sources of income together (eg surveys). Another reason is the ease of translating between Bayesian probabilities and decisions that incorporate costs, rather than scientific studies where you might want to just test a hypothesis. 

I'm a scientist that uses and advocates for Bayesian statistics where appropriate! 

I think it's incorrect to frame it as Bayesian vs Frequentist (as someone who has TA'ed and taught Bayesian stats courses) in general. It really does depend on the context and what you want to do. But, I find that many astronomers tend to use frequentist statistics incorrectly to test a hypothesis and that perhaps is where some of the "adversarial" mindset of B vs F comes in. 

[bayessays]

24 minutes ago, TakeruK said:

I'm a scientist that uses and advocates for Bayesian statistics where appropriate! 

I think it's incorrect to frame it as Bayesian vs Frequentist (as someone who has TA'ed and taught Bayesian stats courses) in general. It really does depend on the context and what you want to do. But, I find that many astronomers tend to use frequentist statistics incorrectly to test a hypothesis and that perhaps is where some of the "adversarial" mindset of B vs F comes in. 

For sure, I was just trying to give some over-simplified reasons. Definitely didn't mean to imply scientists don't use Bayesian methods, and definitely agree that the adversarial mindset is not useful. It really depends on the problem you want to solve.

[Bayesian1701]

@Zaphod2020 Sorry but  because it is an odd problem that would make it very easy to identify me through a google search if you knew what I did.  I would like be remain anonymous here for now at least.  

Edit to add:  Also I hate the Bayesian is better always debate because I don’t think it is always better.  Maybe that’s because I have done research with a frequentist for three years and been exclusively taught by frequentists.   Sometimes Bayesian statistics is better and sometimes it over complicates thing and the prior beliefs can screw up the results.  

Edited March 7, 2018 by Bayesian1701

[StatsG0d]

18 hours ago, Zaphod2020 said:

Reviewing the research interests of professors across various departments, I've noticed that statisticians who apply their work in the social sciences tend to be Bayesians. Do social scientific problems lend themselves to Bayesian methods or am I just making this up?

To me, I don't think this is true. I think it's just that statisticians have a competitive advantage in applying Bayesian methods to social sciences (since they don't usually teach such methods in social sciences). For example, economists, which are arguably the most mathematically inclined social scientists, rely almost exclusively on frequentist regression (albeit at a very high level). So since statisticians are more familiar with Bayesian methods than economists, the former have a comparative advantage in applying Bayesian methods to social sciences.

[Zaphod2020]

6 hours ago, footballman2399 said:

To me, I don't think this is true. I think it's just that statisticians have a competitive advantage in applying Bayesian methods to social sciences (since they don't usually teach such methods in social sciences). For example, economists, which are arguably the most mathematically inclined social scientists, rely almost exclusively on frequentist regression (albeit at a very high level). So since statisticians are more familiar with Bayesian methods than economists, the former have a comparative advantage in applying Bayesian methods to social sciences.

That's a really interesting insight.