I work as a healthcare actuary. Three years ago, I took a Coursera class on Bayesian Statistics (UCSC Matthew Heiner) and learned rjags, and a year after that, I did the Statistical Rethinking lectures on YouTube and learned some Stan.
Most of my day-to-day model building involves summarizing data in SAS and displaying those means in Excel with a few calculations. I sometimes do a side project in R and Stan (cmdstanr, brms, rethinking) to learn by doing. What are the chances, fellow Stan users, you know of actuarial companies or positions that use Bayesian inference as the core modeling approach?
I checked the Jobs section and didn’t see any obvious fit. My background is in healthcare, so some of the pharmacology posts looked interesting. That said, bonus question: do employers expect PhDs to fill the types of positions that use Stan? I have a BS in Actuarial Science, and I’m a Fellow of the Society of Actuaries (US), but that may not hold the same weight outside of the actuarial field as it does inside it.
You might want to try and contact the owner of https://www.magesblog.com/; re: your post title, he is an actuary that uses Stan extensively. Not sure he has job ideas, but that blog and associated publications are a good resource on technical applications.
Nope. In fact, for a lot of practical jobs, Ph.D.s are a red flag, because a lot of Ph.D. holders are very theory oriented. I had a very hard time moving from academia to industry convincing people I really wanted to work on practical problems.
I built most of the code for Stan version 1 with Daniel Lee, who didn’t have a Ph.D. The two software engineers working with me now don’t have Ph.D.s. Most of the people working in sports analytics using Stan don’t have Ph.D.s.
It’s not Stan, but I know PyMC Labs are looking to hire people and I don’t think they have a Ph.D.-only bias.
I “only” have 2 masters degrees, no PhD. I started at a large casualty insurance company in pricing (rate making, loss triangles, etc.). Also mostly in SAS, no Bayesian stuff at the time. I became frustrated with actuarial stuff because new methods weren’t as valued due to the regulatory environment insurance companies are subjected to.
I moved to work in marketing and someone at the company built a small program in a language I had never heard about, called Stan. This was 2016-2017. I was shocked that I could build the model I wanted with the constraints I had in place without having to write a ton of custom code. Many, many attempts at getting some crazy models to compile got me to a good, though basic, understanding of Stan. @Bob_Carpenter even gave me a few pointers at the beginning just as he still does for newcomers today! I continued to work through Stan models and helping people on the forums with their problems gave me an even better understanding.
Today, I have various personal research projects that I’m working on with various people across the globe while I continue to do my day job (albeit at a different marketing company from the one I started at). I have a few preprints on arxiv and hopefully many more to come. Who said you have to have a PhD to do research? Many opportunities have opened up because people see what I’ve done. Almost no one cares about pedigree after a few years out (you don’t want to work somewhere where they have this barrier anyway) it’s all about what you’ve done and what you can do.
Hi @franzsf , I spoke to Markus (for it is he) just last week about this very issue. Yes, there are Bayesians employed in insurance in London. That’s as much as we know, really. He hosts a quarterly meetup and it gets a good turnout. My own experience also suggests that the healthcare - insurance boundary (if that’s where you’re looking) is increasingly multi-faceted, with consulting, private providers, (re)insurance, and all manner of weird healthtech/IoT/wearables startups. Now, each of those, it seems, starts with the usual graduates who know scikit-learn and ChatGPT. Then, some of those companies move up and up over time. But it takes a looong time before they start thinking Bayes. So choose carefully and don’t be the only Bayesian in the room. In some other sectors, Bayes is more common from the get go.
Thanks, @robertgrant. I work in health policy, and find Bayes extremely useful; uncertainty matters (or should matter) to decisionmakers as much as point estimates. But haven’t found it used much either outside the clincal side either.
@franzsf, Thanks for pointing me to www.magesblog.com. I spied some Statistical Rethinking influences there, like the lynx and hare time series. It’s cool to see some Bayes examples applied to insurance concepts.
@bob_carpenter, @spinkney, thank you for the information. I had anecdotal stories of friends who didn’t get into certain jobs because they didn’t have a PhD, and I wasn’t sure how common that was in non-academia Bayes-related industries. For actuaries, it’s mostly about whether you have your credentials or prior experience, regardless of degree (I know a few former entomologists!).
@robertgrant, thank you for the observations. I am the only Bayesian in the room where I’m currently at, ha! I am working on converts but also trying to benchmark my skill in the topic and tools, so I’m not leading anyone astray accidentally.
@spinkney, how did you come across your collaboration projects? I write about Bayes-related topics on my website. Still, I feel some imposture syndrome because I don’t have any Bayes modeling that’s gone through peer review (e.g., an internal quality control process or formal research process).
We’re all impostures at some level so don’t worry about it! I don’t have any papers that have gone through peer review. Luckily most of my stuff has the quality of the “proof is in the pudding”. Meaning, any other computational statistician can literally tell whether it’s correct or not just by running the method or program in a PPL. I try to be pretty open about the outstanding issues with the methods. Peer review would catch that - if I didn’t - and, ideally, make the method or isssues easier to understand by having me flesh out more details or add additional citations that I didn’t see.
Collaboration comes with just doing stuff, posting it online, and attending conferences. I didn’t seek this out. More like people came to me saying that such-and-such method partially solves or helps them and whether I’d be interested in expanding the method(s). There are tons of cool people out there trying their best. I find that the type of people drawn toward Bayesian modeling are the ones who seek out clarity and openness and are generally cool people to work with.
That was just from the first page of Google hits for Query [stan bayesian insurance actuarial].
Peer review is a very noisy process. I got stuff through peer review before I really understood what I was doing properly and I get stuff rejected now where I’m convinced I know more than the reviewers.
No, but you have a lot of code and documentation that’s gone through peer review!
Thanks for the comments and links. I had a chance to talk to a property and casualty (P&C) actuary at Oliver Wyman, who was using R and Stan to fit loss curves and measure parameter risk.
This thread also inspired me to write about some Stan modeling in an actuarial context on my website as a type of Stan portfolio:
The target audience is analysts or actuaries with some statistical background who know little about Bayesian inference. It won’t be a detailed step-by-step R Markdown presentation. I was trying to balance being accessible to newcomers; still, with enough detail, someone with a stronger Bayesian background could assess the reasonableness of the work. If you have feedback, I’m open to ideas.
Thanks for posting this; coming from a 99.5% FFS state, I’ve always had a morbid fascination with MCO operations … though I’m always reminded of those viral "ghost moose"pictures ;)
More seriously, have you considered extended this to look at potential causal system effects that might affect MCO margin – likely embedded in the individual state chararcteristics? MCO competition, provider density, covered population composition (kids, ABAWDS, SSI, LTSS?), etc. Perhaps more economic than actuarial, but just a random suggestion.
@franzsf
I agree with your initial structural causal model and that the way states design their Medicaid programs plays a role in MCO underwriting outcomes (means and standard deviations):
Pharmacy included in managed care vs. non-risk
ACA Expansion state
Standalone MCOs for Behavioral Health or Long-term Services and Supports (LTSS) vs. integrated
Risk mitigation arrangments, risk adjustment (e.g., CDPS, CRGs), or both
State population, Medicaid size split across X number of MCOs
Underwriting margin component forecasted in capitation rates
Restriction on for-profit MCOs
State eligibility policies
So forth…
This data did have some but not all of these variables. I used year, state, and MCO as a proxy for these considerations to minimize the standard deviation picking up between-population variance—between years, states, and MCOs.
A Ph.D. is a kind of certification. You don’t need any kind of degree to do research, but you might need it to satisfy the HR requirements at some jobs. Professorships are typically only offered to people with Ph.D.s. I don’t know that we’d hire a research scientist here without a Ph.D., but there’s no hard-and-fast rule. You’ll have a hard time getting past HR without an undergraduate degree, but it’s not impossible (I know folks who work at high-level software engineering jobs at places like Amazon who dropped out of their undergrad programs). I had a colleague at Bell Labs in signal processing who also never finished his undergraduate degree. The two best programmers I’ve ever met dropped out of their undergraduate programs. I work with a ton of people who do research without a Ph.D. I’m usually surprised they don’t have a Ph.D. because most (but obviously not all) people who are into research are also into school.