Statistics PhD options: Berkeley v.s. Harvard v.s. Duke v.s. Columbia

[DanielWarlock]

I have currently received offer from Duke, Berkeley and Harvard. The only school I will consider over these but have not heard from now is Columbia because of geographic proximity (my family is located at NYC).

I would deeply appreciate opinions and advice on choosing between these 4 programs. The following is my situation and concerns:

Research interest: computational statistics and applied probability. I'm not dead set on any particular topics at this point. Currently I work on Monte Carlo (MCMC, SMC etc.) But I would be interested to work with Bayesian, non-parametrics, or applied probability such as random matrices, stochastic process. More applied topics such as machine learning, deep learning, high-dimensional statistics would be appealing to me as well. I'm not dead set to go into academia; research positions in industry (tech companies such as Microsoft research, Google Deep mind etc.) are appealing to me as well but it seems that they tend to go for people with a particular types of skill set. For example, an expert in MCMC may not be particularly suitable for industry positions. 

Other academic concerns: How difficult/stressful is the curriculum and qual exams in each department? And how stressful in general is each department? I know Harvard is more laid back but not sure about the others. I'd prefer to go to a place where people are not stressed out all the time--that's why I didn't apply to Uchicago and Stanford in the first place. If someone can comment on department culture at Columbia, it would be particularly useful.

Personal: I'm based in NYC with my wife and do not like to go too far away to the west coast and this is why I hesitate choosing Berkeley despite the research interests there seem most diverse. This is also why I would strongly consider Columbia. 

Note: I will go to visit days for sure. But it is good to hear opinions on this forum as well. 

 

 

 

 

[Stat Assistant Professor]

Congratulations on your acceptances! Those are some fine places. I cannot comment so much on the cultures of the departments, but for what it's worth:

Columbia has David Blei who works on Bayesian nonparametrics and computational statistics (mainly from an optimization-based approach, e.g. variational inference). His students and postdocs are quite successful at securing jobs in both academia and industry (e.g. Google, Apple, Deep Mind). The same is true of Michael Jordan at UC Berkeley who was Blei's PhD advisor. Students and postdocs of Jordan do extremely well in the job market.

Harvard strikes me as a big MCMC department. It seems to me that in the machine learning community (rather than the "pure statistics" community), variational inference methods are currently of greater interest than MCMC. But there are also people working on scalable MCMC so that it can be more attractive to "big data" practitioners (possibly some at Harvard). I know that at Duke, there are also several people working on things like approximate MCMC and "embarrassingly parallel" MCMC (e.g. David Dunson). Duke also seems to have the shortest PhD completion time on average (most students seem to finish in four years), if that is an important factor to you.

Overall, I would say that you should definitely consider things like geography and quality of life. If you choose to go the academic route, you may not have much of a choice in where you end up geographically, unless you are a superstar (and even then, it's not a "sure" thing). So the PhD may potentially be the only time that you have an enormous say in where exactly you want to go.

[DanielWarlock]

45 minutes ago, Stat PhD Now Postdoc said:

Congratulations on your acceptances! Those are some fine places. I cannot comment so much on the cultures of the departments, but for what it's worth:

Columbia has David Blei who works on Bayesian nonparametrics and computational statistics (mainly from an optimization-based approach, e.g. variational inference). His students and postdocs are quite successful at securing jobs in both academia and industry (e.g. Google, Apple, Deep Mind). The same is true of Michael Jordan at UC Berkeley who was Blei's PhD advisor. Students and postdocs of Jordan do extremely well in the job market.

Harvard strikes me as a big MCMC department. It seems to me that in the machine learning community (rather than the "pure statistics" community), variational inference methods are currently of greater interest than MCMC. But there are also people working on scalable MCMC so that it can be more attractive to "big data" practitioners (possibly some at Harvard). I know that at Duke, there are also several people working on things like approximate MCMC and "embarrassingly parallel" MCMC (e.g. David Dunson). Duke also seems to have the shortest PhD completion time on average (most students seem to finish in four years), if that is an important factor to you.

Overall, I would say that you should definitely consider things like geography and quality of life. If you choose to go the academic route, you may not have much of a choice in where you end up geographically, unless you are a superstar (and even then, it's not a "sure" thing). So the PhD may potentially be the only time that you have an enormous say in where exactly you want to go.

 

Thank you so much for the reply. I have heard of Blei and Jordan of course. Andrew Gelman seems to be another prominent Bayesian at Columbia. But do I have a sure chance of securing these people as my supervisor though? I know at Harvard PhD students can pretty much talk to and work with any faculty they want and there is flexibility--a primary supervisor is not set until like the 3rd year. But how does phd supervisor match typically go at other places?

[Stat Assistant Professor]

It would be a good idea to ask that during your visit days... though if there are any lurkers on this forum who currently attend those programs, maybe they can answer.

[Geococcyx]

From Duke's visit day, their situation seems similar to what you've heard about Harvard.  A recent poster suggested that some professors may have preferences for students with stronger math backgrounds, but I'm guessing you wouldn't be at much risk of missing out.

Edited February 12, 2020 by Geococcyx

[icantdoalgebra]

I don't know how difficult the qualifying or prelim exams are at Berkeley, but the introductory courses for PhD students range from being fairly challenging to insanely difficult. Measure theoretic probability, taught by Jim Pitman, is very difficult (but I've also heard very rewarding) to the point where even PhD students drop out of the class. The introductory stats classes also aren't the standard Casella and Berger, Stat 210A is very similar to the content covered in it, but Stat 210B is essentially a course on modern high-dimensional statistical techniques. However most of the homework problems (especially near the end of the course) are rederiving fairly influential theoretical results from the last decade or so (you can look up the papers they came from but that defeats the purpose of trying to solve them).

However from what I've seen, the grad students here don't seem too stressed out; the area around UC Berkeley is not the safest but there are a lot of interesting things, unique to California and the Bay Area, to do which you won't find on the East Coast. Many of them actually seem to enjoy what they're doing, although from what I've heard, the rent does eat up quite a bit of the living stipend, especially if you want to live near campus (although NYC isn't that different in that regards). 

 

Edited February 13, 2020 by icantdoalgebra

[insert_name_here]

2 hours ago, icantdoalgebra said:

I don't know how difficult the qualifying or prelim exams are at Berkeley, but the introductory courses for PhD students range from being fairly challenging to insanely difficult. Measure theoretic probability, taught by Jim Pitman, is very difficult (but I've also heard very rewarding) to the point where even PhD students drop out of the class. The introductory stats classes also aren't the standard Casella and Berger, Stat 210A is very similar to the content covered in it, but Stat 210B is essentially a course on modern high-dimensional statistical techniques. However most of the homework problems (especially near the end of the course) are rederiving fairly influential theoretical results from the last decade or so (you can look up the papers they came from but that defeats the purpose of trying to solve them).

However from what I've seen, the grad students here don't seem too stressed out; the area around UC Berkeley is not the safest but there are a lot of interesting things, unique to California and the Bay Area, to do which you won't find on the East Coast. Many of them actually seem to enjoy what they're doing, although from what I've heard, the rent does eat up quite a bit of the living stipend, especially if you want to live near campus (although NYC isn't that different in that regards). 

 

I graduated from Berkeley stats PhD - the courses are rigorous, but not crazy. I haven't really heard of PhD students dropping classes because they couldn't handle it.

We do have a pretty laid back set of requirements - no written qualifying exams, only an oral exam you take sometime between your second and fifth year which students never fail.

Anecdotally, I've heard Gelman is very difficult to work with (a Columbia PhD student volunteered that on my visit day)

Berkeley is really great, but I wouldn't stress too much - you can get a great education at any of those schools. Just find somewhere that you'll be happy!

[bayessays]

Obviously I don't think you can go wrong here.  On your visits, definitely figure out the ease of working with top faculty.  I think Duke probably has an advantage just in terms of the sheer number of elite faculty members working in similar areas. At Duke, you have Dunson, Reiter, Hoff, Gelfand, West, Mukherjee and more- I see these people's papers and former students everywhere in top departments. And since Duke has such a Bayesian/computational focus, I think that might maximize your chance of working with top faculty who matches your interests.  This is in contrast to Berkeley or Columbia which have Jordan/Blei, but then not necessarily huge cores of faculty in similar areas.

Berkeley and Harvard also have lots of top faculty but in many different areas.  Harvard has Meng, Murphy, Imai, Kou -- all great, but working in totally different areas.  Berkeley has van der Laan, Wainwright, but again, totally different areas that you might not be interested in.  There is more uncertainty there about what problems you'd be able to work on.

I think Columbia, besides Blei/Gelman, is probably a step down from the others, but you can dig into specific faculty more to see if they interest you -- Peter Orbanz has similar interests to you, but he seems to have left for UCL recently.

You seem to not want to go to Berkeley for family reasons, so I don't think there's any reason you should completely uproot your life to go there.  It's not going to give you some big advantage over the other schools.  If staying in NYC is important, I think Columbia would be a reasonable choice too.

I don't know your financial situation, but Duke is probably going to be the only place there where your stipend will be able to support two people, as Durham is half the cost of the other 3 locations.

[mikstabio]

Agree with above that Duke is also a good choice (I was a BS+MS statsci student there). Although from the profs bayessays mentions, Reiter barely does Bayesian stats or ML(he is more into missing data etc. and is quite non-theoretical). Gelfand is great if spatial stats interests you, but he is at an almost-retiring stage. Dunson, Hoff, West, Mukherjee are all excellent choices; even Tokdar (for bayes theory) and Li Ma (more computational focused), I'll add.

[Stat Assistant Professor]

Having been on both the primary (TT) and secondary (postdoc/VAP) academic job markets twice now, I will reiterate that if there is any chance that you are considering an academic job, you should definitely prioritize quality of life for your PhD, because there won't be as many opportunities to be geographically mobile afterwards. The academic job market in Statistics/Biostatistics is crowded and competitive enough these days that many PhD holders/postdocs from the Ivy League, Stanford, Berkeley, etc. are taking jobs at schools whose programs are ranked much lower than their alma mater/postdoc institution. If you can publish extremely prolifically during your PhD and postdoc *in the top journals* (a long publication list without -any- in top venues typically won't cut it either), you can certainly improve your chances of landing at a "dream" school in your most desired location. But that isn't guaranteed.

I decided that the academic lifestyle suits me and that the job security and relative freedom in day-to-day job duties were worth sacrificing the geographical flexibility and taking a slightly lower salary than in industry. But I could see that this is (understandably) not the case for everyone.

Edited February 13, 2020 by Stat PhD Now Postdoc

[MathStat]

btw, Gelfand from Duke retired 2 or 3 years ago. So did Robert Wolpert last year. At Duke there is also Jonathan Mattingly, from math (joint with stats) who is very interested in MCMC and who's a pretty cool guy. 

 I just finished my undergrad at Duke (math major) and I'm now in my first year at UChicago. Last year I was also admitted to all the schools you mentioned in the title, except Duke who told me they wanted me to go somewhere else. I can confirm that Duke is an absolutely wonderful place, where professors are excellent and they will help you get research straight away and publish a lot. All the names mentioned above are absolutely phenomenal choices. 

A word about berkeley: turning it down may be the hardest thing ever and you may look back from time to time...so I would say be absolutely 100% certain you definitely don't wanna go there. 

However Duke has better living costs, comparable weather and I would say that anything is possible in terms of academic and research in industry job placements: for instance natesh Pillai who worked with Wolpert and mukherjee is now a top prof at Harvard, James Johndrow who worked with Dunson (and several other ppl) is now a prof. at Wharton etc...there were also a couple students who got jobs at Facebook or Microsoft research. 

I can't comment that much about Harvard and Columbia as I was not that interested in the research going on there. Obviously both are also fantastic choices where the sky seems to be the limit of what you can achieve, but perhaps they're not as friendly and chill as Duke from what I heard.