Best PhD programs for Causal Inference

[forthorn]

I am applying to stat/biostat PhD programs this fall and looking for programs where I could work on causal inference. I know Harvard biostats and Berkeley are good for this, but where else would be good to look into?

[bayessays]

Hopkins has Stuart, UNC has Kosorok, Michigan has a few people working on that stuff in both biostat and stats.  Washington just hired one of Judea Pearle's students. UPenn has quite a few people working in causal inference and might be the closest thing to a department that has a "focus" on it, although the number of people working on causal inference has grown *a lot* in the past few years.  I think you'll find at least one person doing it in most top 10 biostat depts, but you'll have to look through faculty pages to see what's actually interesting to you.

 

[StatsG0d]

On 5/22/2021 at 2:04 PM, bayessays said:

Hopkins has Stuart, UNC has Kosorok, Michigan has a few people working on that stuff in both biostat and stats.  Washington just hired one of Judea Pearle's students. UPenn has quite a few people working in causal inference and might be the closest thing to a department that has a "focus" on it, although the number of people working on causal inference has grown *a lot* in the past few years.  I think you'll find at least one person doing it in most top 10 biostat depts, but you'll have to look through faculty pages to see what's actually interesting to you.

 

Kosorok really works more in precision medicine, which is kind of a special case of causal inference. Hudgens works in causal inference with a focus on interference (i.e., when one individual receiving treatment impacts the probability that another individual receives treatment).

A lot of faculty in the Epi department at UNC works in causal inference theory (e.g., Stephen Cole). 

[StatsG0d]

On 5/22/2021 at 1:17 PM, forthorn said:

I am applying to stat/biostat PhD programs this fall and looking for programs where I could work on causal inference. I know Harvard biostats and Berkeley are good for this, but where else would be good to look into?

If you have the prerequisites, epi and econometrics are good disciplines for causal inference research as well.

[cyberwulf]

5 hours ago, StatsG0d said:

If you have the prerequisites, epi and econometrics are good disciplines for causal inference research as well.

I wouldn't recommend that someone go to an Epidemiology Ph.D. program with the intention of doing methodological research in causal inference. Outside of Harvard (and to a lesser extent UNC), I can't think of any Epi programs with multiple faculty doing causal methods work. 

[bayessays]

5 hours ago, StatsG0d said:

Kosorok really works more in precision medicine, which is kind of a special case of causal inference.

I'd love to understand the distinction you're making here.  Isn't everybody's research in some sense specialized?  If Kosorok's recent papers have "causal" in the title and he's using, for instance, a potential outcomes framework to estimate causal effects in some new precision medicine setup, is that very different from what most people working on "causal inference" will be doing? I know there are some different schools of thought on the causal inference subject, but isn't Kosorok's work pretty much what you'll get at most places for a person researching causal inference? Or are you comparing him to someone like Judea Pearle or Tyler VanderWeele who are I suppose researching more "foundational" causal questions rather than new applications?  I feel like there aren't a ton of people doing the latter, but would love to hear your thoughts. Thanks!

[StatsG0d]

1 hour ago, bayessays said:

I'd love to understand the distinction you're making here.  Isn't everybody's research in some sense specialized?  If Kosorok's recent papers have "causal" in the title and he's using, for instance, a potential outcomes framework to estimate causal effects in some new precision medicine setup, is that very different from what most people working on "causal inference" will be doing? I know there are some different schools of thought on the causal inference subject, but isn't Kosorok's work pretty much what you'll get at most places for a person researching causal inference? Or are you comparing him to someone like Judea Pearle or Tyler VanderWeele who are I suppose researching more "foundational" causal questions rather than new applications?  I feel like there aren't a ton of people doing the latter, but would love to hear your thoughts. Thanks!

It's a fair question. To me, most of causal inference is concerned with identifying a population average treatment effect (typically not adjusted for covariates), while precision medicine is mostly concerned with which treatment for which individual at what time. Most of traditional causal inference utilizes classical statistical techniques (e.g., regression, GLM, etc.), albeit with some adjustments to account for confounding. In causal inference, it's really important to prove things such as consistency and asymptotic normality.

In precision medicine, a lot of the methods are more machine learning focused. They might prove consistency, but asymptotic normality is a bit rarer.

I guess I just feel precision medicine, while a specific case of causal inference, is a lot different than other fields with specific cases.

[trynagetby]

Harvard, Berkley, UW Stats all have at least 3 very top/rising star type people doing causal inference research. An important distinction you have to make is whether you want to do "classical" causal inference  (propensity scores, average treatment effects, instrumental variables, potential outcomes framework) or "modern" causal inference (dags,judea pearl causal discovery,  reinforcement learning, adaptive designs etc...). Both are pretty hot right now but the flavor of research is extremely different.

 

[cyberwulf]

On 5/24/2021 at 9:18 PM, trynagetby said:

Harvard, Berkley, UW Stats all have at least 3 very top/rising star type people doing causal inference research. An important distinction you have to make is whether you want to do "classical" causal inference  (propensity scores, average treatment effects, instrumental variables, potential outcomes framework) or "modern" causal inference (dags,judea pearl causal discovery,  reinforcement learning, adaptive designs etc...). Both are pretty hot right now but the flavor of research is extremely different.

 

This is a great answer. Another way of thinking about the distinction is that there is theoretical (= "classical") causal inference, which is about defining and exploring the properties of new estimands and identification approaches, and there is methodological/applied (= "modern") causal inference, which uses the potential outcomes framework and techniques from causal inference to answer specific scientific questions. I tend to think of the former when I think of the term "causal inference"; the latter includes several areas that don't carry the causal name: precision medicine, dynamic treatment regimes, treatment effect heterogeneity, etc.

[MeltedSnowMan]

If you look at the ongoing research, a lot of interest and focus has been recently dedicated to causal inference in high dimensional settings. Top authors in this field at the moment are Viktor Chernozhukov (MIT), Athey and Wager (Stanford), Alexandre Belloni (Duke). Athey's husband, Guido Imbens, has also written one of the best book out there in causal inference (published with Rubin, who developed the potential outcome framework) and is also at Stanford, so I would definitely encourage you to apply to their econ phd. Not only this research is rather fresh and tries to leverage modern machine learning techniques, it also lacks a lot of empirical applications using those techniques and there is still a lot of ground for extending the current theoretical results (that's what these authors are actually doing atm). Berkeley is a good choice too, they recently hired one of Chernozhukov's students I believe. 

[StatsG0d]

On 5/24/2021 at 9:18 PM, trynagetby said:

Harvard, Berkley, UW Stats all have at least 3 very top/rising star type people doing causal inference research. An important distinction you have to make is whether you want to do "classical" causal inference  (propensity scores, average treatment effects, instrumental variables, potential outcomes framework) or "modern" causal inference (dags,judea pearl causal discovery,  reinforcement learning, adaptive designs etc...). Both are pretty hot right now but the flavor of research is extremely different.

 

This is exactly what I was trying to say regarding Kosorok vs. Hudgens and causal inference / precision medicine above, but is a more elegant and general answer. Couldn't agree more.

On 5/24/2021 at 4:48 PM, cyberwulf said:

I wouldn't recommend that someone go to an Epidemiology Ph.D. program with the intention of doing methodological research in causal inference. Outside of Harvard (and to a lesser extent UNC), I can't think of any Epi programs with multiple faculty doing causal methods work. 

I totally agree with you, but I figured I'd let the OP decide / figure out which programs are suitable.