slimp estimates linear models using Stan and Pandas. Think rstanarm or brms, but in Python and faster.
Create the model:
import matplotlib.pyplot
import numpy
import pandas
import slimp
y, x = numpy.mgrid[0:10, 0:10]
z = 10 + x + 2*y + numpy.random.normal(0, 2, (10, 10))
data = pandas.DataFrame({"x": x.ravel(), "y": y.ravel(), "z": z.ravel()})
model = slimp.Model("z ~ 1 + x + y", data, num_chains=4)
# Also possible to specify random seed
# model = slimp.Model("z ~ 1 + x + y", data, seed=42)
Sample the parameters, check the results:
model.sample()
print(model.hmc_diagnostics)
print(model.summary()[["N_Eff", "R_hat"]].describe().loc[["min", "max"], :])
r_squared = slimp.r_squared(model)
print(r_squared.quantile([0.05, 0.95]))
Plot prior and posterior predictive checks:
figure, plots = matplotlib.pyplot.subplots(1, 2, layout="tight", figsize=(8, 4))
slimp.predictive_plot(model, use_prior=True, plot_kwargs={"ax":plots[0]})
slimp.predictive_plot(model, use_prior=False, plot_kwargs={"ax":plots[1]})
Plot the credible intervals of the parameters and their distributions:
slimp.parameters_plot(model, include=["x", "y"])
slimp.KDEPlot(model.draws["sigma"], prob=0.90)
Use a custom Stan model: have a look here