When HDI and Equal-Tailed Intervals Differ
What you'll learn: You'll discover exactly when Highest Density Intervals (HDI) and equal-tailed credible intervals give you different answers, and why it matters.
The Core Distinction
While both HDI and equal-tailed intervals capture 95% (or any chosen percentage) of your posterior distribution, they can diverge dramatically when your posterior isn't symmetric. Equal-tailed intervals always put 2.5% probability in each tail. HDI always includes the most probable values—the densest region.
When They Differ
Skewed Posteriors
Imagine you're estimating a proportion with limited data and a weak prior. If your posterior is heavily skewed (say, bunched up near 0.8 with a long tail stretching down to 0.3), the equal-tailed interval will dutifully exclude 2.5% from each end—even though the left tail is sparse and the right side is densely packed.
The HDI, however, cuts off the least-dense regions first. It might exclude all of that sparse left tail and trim a bit from the right, giving you a narrower, more informative interval that contains the peak.
Analogy: Think of a crowded concert where everyone's packed near the stage (right side) but a few people wander far back (left tail). Equal-tailed splits the crowd evenly by percentile. HDI keeps everyone close to the action.
Multimodal Posteriors
When your posterior has two or more peaks (common when combining conflicting data sources), HDI can even become discontinuous—two separate intervals around each peak. Equal-tailed intervals always give one continuous range, potentially including low-probability "valleys" between modes.
Practical Impact
HDI is generally more interpretable for decision-making because it focuses on where the mass actually is. Equal-tailed intervals are easier to compute but can be misleading with asymmetric distributions.
Key Takeaway: HDI and equal-tailed intervals diverge meaningfully when posteriors are skewed or multimodal—HDI captures the densest credible regions, while equal-tailed intervals mechanically split tail probabilities regardless of shape.