Ordinal Regression Prior

numpyro
1

In the notebook for ordinal regression, model3 has a peculiar prior, I think.

If I simulate from it, I get this skewed distribution due to the OrderedTransform:

import numpyro
import numpyro.distributions as dist
import jax
import matplotlib.pyplot as plt

rng_key = jax.random.key(42)
x = dist.TransformedDistribution(dist.Normal(0, 1).expand([3]), dist.transforms.OrderedTransform()).sample(rng_key, sample_shape=(10000,))

plt.hist(x, 100, histtype="step", fill=False,)

Prior histogram model3
Prior histogram model3561×413 9.03 KB

I feel like this was not intended. Maybe we should replace it with sampling from the Normal and then ordering?

y = dist.Normal(0,1).expand([3]).sample(rng_key, sample_shape=(10000,)).sort()

plt.hist(y, 100, histtype="step", fill=False,)

Proposed prior
Proposed prior552×413 10.3 KB

Possibly this is equivalent to model2?