如何在JAGS中通过均值与标准差重新参数化Beta分布？附代码验证

Great question! Your core approach to reparameterizing the Beta distribution using mean (mu) and precision (phi = alpha + beta) is totally correct—those formulas you used:

• alpha <- mu * phi
• beta <- (1 - mu) * phi
  are exactly the right way to map mean/precision to the standard Beta parameters. That said, there are a few clarifications and tweaks needed for your code to work as intended:

1. Fix the Observed Data Logic #

Your comment says theta[i] ~ dbeta(alpha, beta) # theta值来自数据, which has a small logical inconsistency:

• If theta is a latent variable you're inferring from human behavior data (e.g., strategy choices), then this line is fine (you're modeling each subject's theta as coming from the Beta distribution).
• But if theta is directly observed data (e.g., you already calculated each subject's strategy X usage proportion), then theta should live in your data block, and you need to model the likelihood of observing those theta values (swap theta[i] with your observed variable name, like theta_obs[i]).

2. Full Working Model Examples #

Let's cover both scenarios to make this concrete:

Scenario 1: Theta is a Latent Variable (Inferred from Choice Data) #

Suppose you have count data: each subject chose strategy X count_x[i] times out of n_trials[i] total tasks. Here's the complete model:

model{
  # Likelihood: Map observed choices to latent theta
  for (i in 1:n_subjects) {
    count_x[i] ~ dbin(theta[i], n_trials[i])  # Binomial likelihood for choice counts
    theta[i] ~ dbeta(alpha, beta)             # Each subject's theta follows the Beta prior
  }
  
  # Reparameterization: Mean + Precision -> Alpha/Beta
  alpha <- mu * phi
  beta <- (1 - mu) * phi
  
  # Priors
  mu ~ dunif(0, 1)          # Non-informative prior for the population mean theta
  phi ~ dgamma(0.01, 0.01)  # Weakly informative prior for precision (avoids over-constraint)
}

Scenario 2: Theta is Directly Observed Data #

If you already have precomputed theta values for each subject (e.g., theta_obs[i]), simplify the model to fit those observations directly:

model{
  # Likelihood: Fit observed theta values to the Beta distribution
  for (i in 1:n_subjects) {
    theta_obs[i] ~ dbeta(alpha, beta)
  }
  
  # Reparameterization
  alpha <- mu * phi
  beta <- (1 - mu) * phi
  
  # Priors
  mu ~ dunif(0, 1)
  phi ~ dgamma(0.01, 0.01)
}

3. Note on Mean + Standard Deviation #

You mentioned the paper used mean and standard deviation for reparameterization. Just to connect the dots: the Beta distribution's standard deviation sigma relates to mu and phi via:
sigma = sqrt( (mu * (1 - mu)) / (phi + 1) )
While you could reparameterize directly using sigma, working with precision (phi = alpha + beta) is more straightforward in JAGS since it maps linearly to the Beta parameters.

4. Small Prior Tweak #

Your original prior phi ~ dgamma(.1,.1) works, but using a weaker prior like dgamma(0.01, 0.01) is often better when data is limited—it lets the data drive the inference more than the prior.

Overall, your core reparameterization logic was perfect—just make sure the data/model alignment matches whether theta is latent or observed, and you're good to go!

内容的提问来源于stack exchange，提问作者ramund