Incorrect results with Poisson regression using bambi?

Question

I am giving bambi (version 0.1.0) a try for a simple Poisson regression model. However the results differ when compared to straight pymc3 or statsmodels implementations, and I cannot seem to figure out how to interpret the coefficients that bambi gives me. The test code is below. Did I specify the model wrong, or should I not rely on the automatic priors of bambi?

import numpy as np
import scipy.stats
import pandas
import patsy
import statsmodels
import pymc3
import bambi

%matplotlib inline

# generate data
num_subjects = 4
mu = [5, 8, 10, 11]
num_samples = [43, 60, 56, 38]

counts = [scipy.stats.poisson.rvs(m,size=n,random_state=m) for m,n in zip(mu,num_samples)]
counts = np.concatenate(counts)
subject = np.repeat(np.arange(num_subjects), num_samples)

df = pandas.DataFrame( np.vstack([subject,counts]).T, columns=['subject','counts'])

# sample means
print( df.groupby('subject').mean() )

# subject 0 = 5.0
# subject 1 = 7.4
# subject 2 = 9.5
# subject 3 = 10.0


# fit with bambi
model_bambi = bambi.Model(df)
result_bambi = model_bambi.fit('counts ~ C(subject)', categorical=['subject'], family='poisson', chains=2)

print(result_bambi.summary(hpd=None, diagnostics=None))

# resulting posterior means:
# Intercept        9.3310 -> ?
# C(subject)[T.1]  3.8171 -> ?
# C(subject)[T.2]  4.4419 -> ?
# C(subject)[T.3]  3.8652 -> ?


# fit directly with pymc3
with pymc3.Model() as model_pymc3:
    pymc3.glm.GLM.from_formula("counts ~ C(subject)", df, family=pymc3.glm.families.Poisson())
    trace = pymc3.sample(2000, njobs=2, tune=500)

pymc3.plot_posterior(trace, varnames=[x for x in trace.varnames if x[:2]!='mu']);

# resulting posterior means:
# Intercept        1.6065 -> mu = 5.0 = exp(1.6065) 
# C(subject)[T.1]  0.3990 -> mu = 7.4 = exp(1.6065+0.3990)
# C(subject)[T.2]  0.6477 -> mu = 9.5 = exp(1.6065+0.6477)
# C(subject)[T.3]  0.6977 -> mu = 10.0 = exp(1.6065+0.6977)


# fit with statsmodels
my, mx = patsy.dmatrices( "counts ~ C(subject)", df, NA_action='raise')
model_sm = statsmodels.api.GLM(my, mx, family=statsmodels.api.families.Poisson())
result_sm = model_sm.fit()

print(result_sm.summary())

# resulting posterior means:
# Intercept        1.6094 -> mu = 5.0 = exp(1.6094) 
# C(subject)[T.1]  0.3965 -> mu = 7.4 = exp(1.6094+0.3965)
# C(subject)[T.2]  0.6456 -> mu = 9.5 = exp(1.6094+0.6456)
# C(subject)[T.3]  0.6958 -> mu = 10.0 = exp(1.6094+0.6958)

score 2 · Accepted Answer · answered Nov 23 '17 at 01:31

My apologies for the (very) slow reply; I wasn't subscribed to the [bambi] tag (but am now), and only just saw this. This is indeed a bug (details are here). I just opened a PR for it, so if you clone from the repo, the problem should be solved (and I'll issue a new PyPI release sortly). I realize this is probably not much use to you at this point, but thanks for flagging it anyway. If you run into any similar issues in future, please open an issue on the GitHub repo, as this definitely falls into bug territory.

Incorrect results with Poisson regression using bambi?

1 Answers1