5 ms·

Bayesian Structural Equation Modeling using blavaan (2022)

58 points by gravitate 3y ago

3abiton 3y ago
I feel the technical barrier of adoption of bayesian methods is still high enough to deter potential users.
- ASpring 3y ago
  It is still the Linux of statistical software. Rstanarm is getting quite good though for R users.
- mike_ivanov 3y ago
  It is much less challenging with Bambi[1] and brms[2]. [1] https://bambinos.github.io/bambi/ [2] https://paul-buerkner.github.io/brms/
  - eynsham 3y ago
    If you put the links in plaintext they're not clickable.
- mnky9800n 3y ago
  I completely agree with you and this is why we are developing 4d-modeller: https://4dmodeller.github.io/fdmr/ https://4dmodeller.github.io/fdmr/ It's still in early stages but the concept is that these methods are mostly opaque even to highly technical users. So we start with shiny apps that help you build a model and that will work up to (not yet implemented, but we will do a sprint in a couple weeks) wrapper functions, which helps you get things going until you start wanting to get more and more complex. We just ran a Hackathon and participants with no R or Bayesian experience were able to make models.
j7ake 3y ago
What’s the advantage of using Stan when now there is numpyro and pymc. The newer frameworks seem much more flexible and performant than Stan.
- nextos 3y ago
  For things that do not require more expressiveness than what Stan offers, I have found the speed and quality of sampling is outstanding. Furthermore, thanks to static typing, complex models are easy to write. To be fair, Pyro is also great and is my go-to for massive models where MCMC is unfeasible or for those requiring more expressiveness and/or deep components. I have less experience with PyMC.
- bigfudge 3y ago
  brms and rstanarm make using it very easy. Blavaan likewise. The transition from Conventional model syntax is great scaffolding.
aquafox 3y ago
FWIW, there's also a fast and simple alternative in R called greta (https://greta-stats.org https://greta-stats.org)
- j7ake 3y ago
  Looks “great-ah”!
bigfudge 3y ago
For someone too lazy to read all the docs, is most/all of the lavaan model syntax implemented?

58 points by gravitate 3y ago

3abiton 3y ago

I feel the technical barrier of adoption of bayesian methods is still high enough to deter potential users.

ASpring 3y ago
It is still the Linux of statistical software. Rstanarm is getting quite good though for R users.
mike_ivanov 3y ago
It is much less challenging with Bambi[1] and brms[2]. [1] https://bambinos.github.io/bambi/ [2] https://paul-buerkner.github.io/brms/
- eynsham 3y ago
  If you put the links in plaintext they're not clickable.
mnky9800n 3y ago
I completely agree with you and this is why we are developing 4d-modeller: https://4dmodeller.github.io/fdmr/ https://4dmodeller.github.io/fdmr/ It's still in early stages but the concept is that these methods are mostly opaque even to highly technical users. So we start with shiny apps that help you build a model and that will work up to (not yet implemented, but we will do a sprint in a couple weeks) wrapper functions, which helps you get things going until you start wanting to get more and more complex. We just ran a Hackathon and participants with no R or Bayesian experience were able to make models.

j7ake 3y ago

What’s the advantage of using Stan when now there is numpyro and pymc. The newer frameworks seem much more flexible and performant than Stan.

nextos 3y ago
For things that do not require more expressiveness than what Stan offers, I have found the speed and quality of sampling is outstanding. Furthermore, thanks to static typing, complex models are easy to write. To be fair, Pyro is also great and is my go-to for massive models where MCMC is unfeasible or for those requiring more expressiveness and/or deep components. I have less experience with PyMC.
bigfudge 3y ago
brms and rstanarm make using it very easy. Blavaan likewise. The transition from Conventional model syntax is great scaffolding.

aquafox 3y ago

FWIW, there's also a fast and simple alternative in R called greta (https://greta-stats.org https://greta-stats.org)

j7ake 3y ago
Looks “great-ah”!

bigfudge 3y ago

For someone too lazy to read all the docs, is most/all of the lavaan model syntax implemented?