XIANG YE / BAYESIAN RESEARCH
Directional
Statistics.
When the data have
a sense of direction.
Angles on a circle. Directions on a sphere. Statistical methods that respect the geometry of the data.
359° and 1° are only 2° apart.
The end of the scale meets the beginning.
Set the sample’s mean resultant length; the number of observations stays fixed.
0° at north; angles increase clockwise.
View the counts and scale
Radius ∝ √count. Outer grid ring: 29 observations. Inner rings: 25%, 50%, and 75% of that count. The count scale adapts to the largest bin. At R̄ = 1 the observation marks overlap.
| Bearing | Count |
|---|---|
| 0.0–22.5° | 6 |
| 22.5–45.0° | 13 |
| 45.0–67.5° | 29 |
| 67.5–90.0° | 29 |
| 90.0–112.5° | 13 |
| 112.5–135.0° | 6 |
| 135.0–157.5° | 4 |
| 157.5–180.0° | 3 |
| 180.0–202.5° | 2 |
| 202.5–225.0° | 1 |
| 225.0–247.5° | 2 |
| 247.5–270.0° | 2 |
| 270.0–292.5° | 1 |
| 292.5–315.0° | 2 |
| 315.0–337.5° | 3 |
| 337.5–360.0° | 4 |
Wind & weather
↗Animal movement
↗Geological orientation
↗Time & phase
01 / UNDERSTAND THE GEOMETRY
A circle has
no natural beginning.
Compass bearings and times of day wrap around. Averaging their numerical labels can be misleading: the arithmetic mean of 359° and 1° is 180°, while their circular mean is 0°.
A rose diagram makes the pattern visible by grouping observations into angular sectors. In the diagram above, larger sector areas mean more observations.
How to read a rose diagramON A STRAIGHT SCALE
The numerical average is 180°.
(359° + 1°) / 2 = 180°
The calculation is correct on a line, but its result points away from both observations.
ON THE CIRCLE
The mean direction is 0°.
02 / BAYESIAN RESEARCH
From directions to models.
Priors for circular models
Penalized complexity priors for the concentration of circular distributions.
Explore the researchRegression with a circular response
Link-adjusted von Mises models for angular outcomes and structured predictors.
Explore the researchRegression with a circular covariate
Joint circular models connecting direction and a linear response through shared latent structure.
Explore the research03 / STAN IMPLEMENTATIONS
The methods, worked through in Stan.
Penalized Complexity Priors for Circular Distributions
Construct and calibrate PC priors for von Mises concentration, with circular-uniform and point-mass base models.
Circular Distributions for Regression Models
Explore link-adjusted von Mises regression, from fixed effects to temporal structure and wind-direction data.
Joint circular models
Explore regression models with circular covariates and learn how to extend them to joint modelling.