Prior Distributions for Normal Means
You'll learn how to express your prior beliefs about an unknown population mean using appropriate probability distributions.
The Core Idea
When working with continuous measurements (heights, temperatures, reaction times), we often assume the data follows a normal distribution. In Bayesian inference, we need to specify a prior distribution for the unknown population mean (μ) before seeing the data.
The most natural choice? Another normal distribution.
Why Normal Priors for Normal Means?
Just as we used a Beta prior for proportions (because proportions live between 0 and 1), we use a normal prior for means because:
- Means can be any real number (positive or negative)
- A normal distribution naturally expresses uncertainty about a continuous parameter
- It creates a conjugate prior relationship with normally-distributed data, making calculations tractable
Specifying Your Prior
You need two parameters:
- Prior mean (μ₀): Your best guess for the population mean before seeing data
- Prior standard deviation (σ₀): How uncertain you are about that guess
Example
Suppose you're measuring average customer spending. You believe the mean is around 15). Your prior would be Normal(μ₀ = 50, σ₀ = 15).
A wide σ₀ expresses weak prior knowledge; a narrow σ₀ expresses strong confidence.
Other Suitable Priors
When you have minimal prior knowledge, you might use:
- Flat (uniform) priors over a reasonable range
- Weakly informative priors with large variance
- Student's t distributions for heavier tails (more robust to outliers)
Key Takeaway: Normal priors are the natural conjugate choice for expressing beliefs about population means, specified by a center point (prior mean) and uncertainty level (prior standard deviation).