# The best of both worlds: Hierarchical Linear Regression in PyMC3¶

Today's blog post is co-written by my student Danne Elbers who is doing her masters thesis with me on computational psychiatry using Bayesian methods. This post also borrows heavily from a Notebook by Chris Fonnesbeck.

The power of Bayesian modelling really clicked for me when I was first introduced to hierarchical modelling. In this blog post we will highlight the advantage of using hierarchical Bayesian modelling as opposed to non-hierarchical Bayesian modelling. This hierachical modelling is especially advantageous when multi-level data is used, making the most of all information available by its 'shrinkage-effect', which will be explained below.

Having multiple sets of measurements comes up all the time, in Psychology for example you test multiple subjects on the same task. You then might want to estimate a model that describes the behavior as a set of parameters relating to mental functioning. Often we are interested in individual differences of these parameters but also assume that subjects share similarities (being human and all). Software from our lab, HDDM, allows hierarchical Bayesian estimation of a widely used decision making model but we will use a more classical example of hierarchical linear regression here to predict radon levels in houses.

This is the 3rd blog post on the topic of Bayesian modeling in PyMC3, see here for the previous two:

## The data set¶

Gelman et al.'s (2007) radon dataset is a classic for hierarchical modeling. In this dataset the amount of the radioactive gas radon has been measured among different households in all county's of several states. Radon gas is known to be the highest cause of lung cancer in non-smokers. It is believed to be more strongly present in households containing a basement and to differ in amount present among types of soil. Here we'll investigate this differences and try to make predictions of radonlevels in different county's based on the county itself and the presence of a basement. In this example we'll look at Minnesota, a state that contains 85 county's in which different measurements are taken, ranging from 2 till 80 measurements per county.

In [2]:
%matplotlib inline

import matplotlib.pyplot as plt
import numpy as np
import pymc3 as pm
import pandas as pd

county_names = data.county.unique()
county_idx = data['county_code'].values

The relevant part of the data we will model looks as follows:

In [3]:
Out[3]:
0 AITKIN 0.832909 1.0
1 AITKIN 0.832909 0.0
2 AITKIN 1.098612 0.0
3 AITKIN 0.095310 0.0
4 ANOKA 1.163151 0.0

As you can see, we have multiple radon measurements (log-converted to be on the real line) in a county and whether the house has a basement (floor == 0) or not (floor == 1). We are interested in whether having a basement increases the radon measured in the house.

## The Models¶

### Pooling of measurements¶

Now you might say: "That's easy! I'll just pool all my data and estimate one big regression to asses the influence of a basement across all counties". In math-speak that model would be:

$$radon_{i, c} = \alpha + \beta*\text{floor}_{i, c} + \epsilon$$

Where $i$ represents the measurement, $c$ the county and floor contains a yes (1) or no (0) to the question if the household has a basement. If you need a refresher on Linear Regressions in PyMC, check out my previous blog post. Critically, we are only estimating one intercept and one slope for all measurements over all counties.

### Separate regressions¶

But what if we are interested whether different counties actually have different relationships (slope) and different base-rates of radon (intercept)? Then you might say "OK then, I'll just estimate $n$ (number of counties) different regresseions -- one for each county". In math-speak that model would be:

$$radon_{i, c} = \alpha_{c} + \beta_{c}*\text{floor}_{i, c} + \epsilon_c$$

Note that we added the subindex $c$ so we are estimating $n$ different $\alpha$s and $\beta$s -- one for each county.

This is the extreme opposite model, where above we assumed all counties are exactly the same, here we are saying that they share no similarities whatsoever which ultimately is also unsatisifying.

### Hierarchical Regression: The best of both worlds¶

Fortunately there is a middle ground to both of these extreme views. Specifically, we may assume that while $\alpha$s and $\beta$s are different for each county, the coefficients all come from a common group distribution:

$$\alpha_{c} \sim \mathcal{N}(\mu_{\alpha}, \sigma_{\alpha}^2)$$ $$\beta_{c} \sim \mathcal{N}(\mu_{\beta}, \sigma_{\beta}^2)$$

We thus assume the intercepts $\alpha$ and slopes $\beta$ to come from a normal distribution centered around their respective group mean $\mu$ with a certain standard deviation $\sigma^2$, the values (or rather posteriors) of which we also estimate. That's why this is called multilevel or hierarchical modeling.

How do we estimate such a complex model with all these parameters you might ask? Well, that's the beauty of Probabilistic Programming -- we just formulate the model we want and press our Inference Button(TM).

Note that the above is not a complete Bayesian model specification as we haven't defined priors or hyperpriors (i.e. priors for the group distribution, $\mu$ and $\sigma$). These will be used in the model implementation below but only distract here.

## Probabilistic Programming¶

### Individual/non-hierarchical model¶

To really highlight the effect of the hierarchical linear regression we'll first estimate the non-hierarchical Bayesian model from above (separate regressions). For each county a new estimate of the parameters is initiated. As we have no prior information on what the intercept or regressions could be we are placing a Normal distribution centered around 0 with a wide standard-deviation. We'll assume the measurements are normally distributed with noise $\epsilon$ on which we place a Half-Cauchy distribution.

In [ ]:
indiv_traces = {}
for county_name in county_names:
# Select subset of data belonging to county
c_data = data.ix[data.county == county_name]
c_data = c_data.reset_index(drop=True)

c_floor_measure = c_data.floor.values

with pm.Model() as individual_model:
# Intercept prior
a = pm.Normal('alpha', mu=0, sd=1)
# Slope prior
b = pm.Normal('beta', mu=0, sd=1)

# Model error prior
eps = pm.HalfCauchy('eps', beta=1)

# Linear model
radon_est = a + b * c_floor_measure

# Data likelihood

# Inference button (TM)!
trace = pm.sample(progressbar=False)

indiv_traces[county_name] = trace

## Hierarchical Model¶

Instead of initiating the parameters separatly, the hierarchical model initiates group parameters that consider the county's not as completely different but as having an underlying similarity. These distributions are subsequently used to influence the distribution of each county's $\alpha$ and $\beta$.

In [16]:
with pm.Model() as hierarchical_model:
# Hyperpriors
mu_a = pm.Normal('mu_alpha', mu=0., sd=1)
sigma_a = pm.HalfCauchy('sigma_alpha', beta=1)
mu_b = pm.Normal('mu_beta', mu=0., sd=1)
sigma_b = pm.HalfCauchy('sigma_beta', beta=1)

# Intercept for each county, distributed around group mean mu_a
a = pm.Normal('alpha', mu=mu_a, sd=sigma_a, shape=len(data.county.unique()))
# Intercept for each county, distributed around group mean mu_a
b = pm.Normal('beta', mu=mu_b, sd=sigma_b, shape=len(data.county.unique()))

# Model error
eps = pm.HalfCauchy('eps', beta=1)

# Expected value
radon_est = a[county_idx] + b[county_idx] * data.floor.values

# Data likelihood
In [17]:
with hierarchical_model:
hierarchical_trace = pm.sample(njobs=4)
Auto-assigning NUTS sampler...
100%|██████████| 1000/1000 [00:07<00:00, 126.84it/s]
/opt/miniconda/lib/python2.7/site-packages/pymc3/step_methods/hmc/nuts.py:451: UserWarning: The acceptance probability in chain 1 does not match the target. It is 0.889237663849, but should be close to 0.8. Try to increase the number of tuning steps.
% (self._chain_id, mean_accept, target_accept))
/opt/miniconda/lib/python2.7/site-packages/pymc3/step_methods/hmc/nuts.py:451: UserWarning: The acceptance probability in chain 2 does not match the target. It is 0.636029263947, but should be close to 0.8. Try to increase the number of tuning steps.
% (self._chain_id, mean_accept, target_accept))
/opt/miniconda/lib/python2.7/site-packages/pymc3/step_methods/hmc/nuts.py:467: UserWarning: Chain 2 contains 2 diverging samples after tuning. If increasing target_accept does not help try to reparameterize.
% (self._chain_id, n_diverging))
/opt/miniconda/lib/python2.7/site-packages/pymc3/step_methods/hmc/nuts.py:467: UserWarning: Chain 3 contains 25 diverging samples after tuning. If increasing target_accept does not help try to reparameterize.
% (self._chain_id, n_diverging))
In [18]:
pm.traceplot(hierarchical_trace);

The marginal posteriors in the left column are highly informative. mu_a tells us the group mean (log) radon levels. mu_b tells us that having no basement decreases radon levels significantly (no mass above zero). We can also see by looking at the marginals for a that there is quite some differences in radon levels between counties; the different widths are related to how much measurements we have per county, the more, the higher our confidence in that parameter estimate.

After writing this blog post I found out that the chains here (which look worse after I just re-ran them) are not properly converged, you can see that best for sigma_beta but also the warnings about "diverging samples" (which are also new in PyMC3). If you want to learn more about the problem and its solution, see my more recent blog post "Why hierarchical models are awesome, tricky, and Bayesian".

## Posterior Predictive Check¶

### The Root Mean Square Deviation¶

To find out which of the models works better we can calculate the Root Mean Square Deviaton (RMSD). This posterior predictive check revolves around recreating the data based on the parameters found at different moments in the chain. The recreated or predicted values are subsequently compared to the real data points, the model that predicts data points closer to the original data is considered the better one. Thus, the lower the RMSD the better.

When computing the RMSD (code not shown) we get the following result:

• individual/non-hierarchical model: 0.13
• hierarchical model: 0.08

As can be seen above the hierarchical model performs a lot better than the non-hierarchical model in predicting the radon values. Following this, we'll plot some examples of county's showing the true radon values, the hierarchial predictions and the non-hierarchical predictions.

In [9]:
selection = ['CASS', 'CROW WING', 'FREEBORN']
fig, axis = plt.subplots(1, 3, figsize=(12, 6), sharey=True, sharex=True)
axis = axis.ravel()
for i, c in enumerate(selection):
c_data = data.ix[data.county == c]
c_data = c_data.reset_index(drop = True)
z = list(c_data['county_code'])[0]

xvals = np.linspace(-0.2, 1.2)
for a_val, b_val in zip(indiv_traces[c]['alpha'][::10], indiv_traces[c]['beta'][::10]):
axis[i].plot(xvals, a_val + b_val * xvals, 'b', alpha=.1)
axis[i].plot(xvals, indiv_traces[c]['alpha'][::10].mean() + indiv_traces[c]['beta'][::10].mean() * xvals,
'b', alpha=1, lw=2., label='individual')
for a_val, b_val in zip(hierarchical_trace['alpha'][::10][z], hierarchical_trace['beta'][::10][z]):
axis[i].plot(xvals, a_val + b_val * xvals, 'g', alpha=.1)
axis[i].plot(xvals, hierarchical_trace['alpha'][::10][z].mean() + hierarchical_trace['beta'][::10][z].mean() * xvals,
'g', alpha=1, lw=2., label='hierarchical')