Skip to main content
Cross Validated

Return to Question

Post Closed as "Duplicate" by Christian Hennig, Shawn Hemelstrand, User1865345
edited body
Source Link
compbiostats
compbiostats
  • 1.6k
  • 1
  • 18
  • 26

Here:

Centering prior distributions on MLE/OLS estimates

I ask about centering priors on MLEs in the context of a logistic regression (in my case with only categorical predictors), which I've seen a lot in papers by my own colleagues (who are likely not Bayesians)

@Ben responds that this is bad practice.

The prior should be selected independently of the data (that is, before observing any data, or a statistical summary thereof).

In a related Stan forumForum post, @FrankHarrell also cautions this, as it could lead to inflation of effective sample sizes relative to actual sample sizes. Instead he suggests using the MLEs as starting conditions for MCMC sampling, which would result in more efficient traversal to the posterior distribution.

I'm wondering what others think. Hoping to get a lively discussion going.

Here:

Centering prior distributions on MLE/OLS estimates

I ask about centering priors on MLEs in the context of a logistic regression (in my case with only categorical predictors), which I've seen a lot in papers by my own colleagues (who are likely not Bayesians)

@Ben responds that this is bad practice.

The prior should be selected independently of the data (that is, before observing any data, or a statistical summary thereof).

In a related Stan forum post, @FrankHarrell also cautions this, as it could lead to inflation of effective sample sizes relative to actual sample sizes. Instead he suggests using the MLEs as starting conditions for MCMC sampling, which would result in more efficient traversal to the posterior distribution.

I'm wondering what others think. Hoping to get a lively discussion going.

Here:

Centering prior distributions on MLE/OLS estimates

I ask about centering priors on MLEs in the context of a logistic regression (in my case with only categorical predictors), which I've seen a lot in papers by my own colleagues (who are likely not Bayesians)

@Ben responds that this is bad practice.

The prior should be selected independently of the data (that is, before observing any data, or a statistical summary thereof).

In a related Stan Forum post, @FrankHarrell also cautions this, as it could lead to inflation of effective sample sizes relative to actual sample sizes. Instead he suggests using the MLEs as starting conditions for MCMC sampling, which would result in more efficient traversal to the posterior distribution.

I'm wondering what others think. Hoping to get a lively discussion going.

Source Link
compbiostats
compbiostats
  • 1.6k
  • 1
  • 18
  • 26

Centering Priors on MLEs vs. Using MLEs as Initial Conditions for MCMC

Here:

Centering prior distributions on MLE/OLS estimates

I ask about centering priors on MLEs in the context of a logistic regression (in my case with only categorical predictors), which I've seen a lot in papers by my own colleagues (who are likely not Bayesians)

@Ben responds that this is bad practice.

The prior should be selected independently of the data (that is, before observing any data, or a statistical summary thereof).

In a related Stan forum post, @FrankHarrell also cautions this, as it could lead to inflation of effective sample sizes relative to actual sample sizes. Instead he suggests using the MLEs as starting conditions for MCMC sampling, which would result in more efficient traversal to the posterior distribution.

I'm wondering what others think. Hoping to get a lively discussion going.