Choosing Informative Priors for Proportions
What you'll learn: How to select beta distribution parameters that encode your prior knowledge about a proportion in Bayesian analysis.
When working with proportions (like conversion rates or success probabilities), the beta distribution is your go-to prior in Bayesian inference. But how do you choose the right parameters α (alpha) and β (beta)?
Translating Prior Knowledge into Parameters
The beta distribution has two parameters that shape your prior beliefs:
- α: Think of this as "prior successes"
- β: Think of this as "prior failures"
The more "data" you encode in your prior (higher α + β), the stronger your prior belief—and the more actual data you'll need to shift it.
Practical Approaches
From prior mean and confidence: If you believe a conversion rate is around 20% but aren't very certain, you might choose α = 2, β = 8 (mean = 2/(2+8) = 0.2, total pseudocounts = 10). For stronger belief, scale up proportionally: α = 20, β = 80.
From domain expertise: If historical data suggests 100 conversions from 500 trials, you could use α = 100, β = 400 as an informative prior that reflects actual experience.
The analogy: Think of α and β as "imaginary observations" you're adding to your analysis. α = 1, β = 1 represents minimal prior knowledge (uniform prior). α = 50, β = 50 means you're adding the equivalent of 100 prior observations centered at 50% success rate.
Calibrating Strength
The sum α + β controls prior strength—how much evidence you need to change your mind. Small sums (α + β ≈ 10) are weakly informative. Large sums (α + β > 100) are strongly informative and should only be used when you have solid prior evidence.
Key Takeaway: Choose beta parameters by translating your prior belief about the proportion's center (via α/(α+β)) and your confidence in that belief (via α+β total strength) into concrete numbers that reflect real or hypothetical prior observations.