Bayesian Point Estimation: Posterior Mean, Median, and Mode
What you'll learn: How to extract a single "best guess" parameter value from your posterior distribution using different measures of central tendency and loss functions.
The Core Idea
After performing Bayesian updating, you have a full posterior distribution describing your uncertainty about a parameter. But sometimes you need to report one number—a point estimate. Unlike frequentist methods that use sample statistics, Bayesian point estimation summarizes the posterior distribution itself.
The three most common Bayesian point estimates are:
Posterior Mean
The posterior mean is the expected value of the parameter given the data. It minimizes squared error loss—the penalty you pay if your estimate is off. Think of it as the "balance point" of the posterior distribution.
If your posterior is symmetric and you care about average performance across many estimates, use the mean.
Posterior Median
The posterior median splits the posterior probability in half: 50% of the probability mass lies below it, 50% above. It minimizes absolute error loss—treating overestimates and underestimates symmetrically but linearly rather than quadratically.
Use the median when you want robustness to skewed posteriors or outliers in the distribution's tail.
Posterior Mode (MAP)
The posterior mode—also called the Maximum A Posteriori (MAP) estimate—is the parameter value with highest posterior density. It minimizes 0-1 loss (all-or-nothing: either you're exactly right or completely wrong).
The mode makes sense when you want the "most plausible" single value, though it can be misleading for multimodal or highly skewed distributions.
Choosing Your Estimate
Your choice depends on your loss function—how you penalize estimation errors. Mean for squared loss, median for absolute loss, mode for 0-1 loss. Each tells a different story about the parameter.
Key Takeaway: Bayesian point estimates—posterior mean, median, and mode—condense the full posterior into a single number, each optimal under different loss functions reflecting how you measure estimation error.