Questo esempio è portato dall'esempio PyMC3 notebook A Primer su metodi bayesiani per Multilevel Modeling
Visualizza su TensorFlow.org | Esegui in Google Colab | Visualizza la fonte su GitHub | Scarica taccuino |
Dipendenze e prerequisiti
Importare
import collections
import os
from six.moves import urllib
import daft as daft
import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
import warnings
import tensorflow.compat.v2 as tf
tf.enable_v2_behavior()
import tensorflow_datasets as tfds
import tensorflow_probability as tfp
tfk = tf.keras
tfkl = tf.keras.layers
tfpl = tfp.layers
tfd = tfp.distributions
tfb = tfp.bijectors
warnings.simplefilter('ignore')
1. Introduzione
In questo CoLab ci sarà adatto ai modelli lineari gerarchici (HLM) di diversi gradi di complessità del modello utilizzando il popolare dataset Radon. Utilizzeremo le primitive TFP e il suo set di strumenti Markov Chain Monte Carlo.
Per adattare meglio i dati, il nostro obiettivo è utilizzare la struttura gerarchica naturale presente nel set di dati. Iniziamo con approcci convenzionali: modelli completamente pool e unpool. Continuiamo con i modelli multilivello: esplorazione di modelli di pooling parziale, predittori a livello di gruppo ed effetti contestuali.
Per un notebook correlato anche il montaggio HLM utilizzando TFP sul set di dati Radon, controlla Linear Mixed-Effect regressione in {Probabilità di TF, R, Stan} .
Se avete domande circa il materiale qui, non esitate a contatto (o partecipare) la probabilità mailing list tensorflow . Siamo felici di aiutare.
2 Panoramica sulla modellazione multilivello
Un primer sui metodi bayesiani per la modellazione multilivello
La modellazione gerarchica o multilivello è una generalizzazione della modellazione di regressione.
I modelli multilivello sono modelli di regressione in cui ai parametri del modello costitutivo vengono date distribuzioni di probabilità. Ciò implica che i parametri del modello possono variare in base al gruppo. Le unità di osservazione sono spesso raggruppate in modo naturale. Il clustering induce dipendenza tra le osservazioni, nonostante il campionamento casuale dei cluster e il campionamento casuale all'interno dei cluster.
Un modello gerarchico è un particolare modello multilivello in cui i parametri sono annidati l'uno nell'altro. Alcune strutture multilivello non sono gerarchiche.
ad es. "paese" e "anno" non sono nidificati, ma possono rappresentare gruppi di parametri separati, ma sovrapposti. Motiveremo questo argomento utilizzando un esempio di epidemiologia ambientale.
Esempio: contaminazione da radon (Gelman e Hill 2006)
Il radon è un gas radioattivo che entra nelle case attraverso i punti di contatto con il suolo. È un cancerogeno che è la causa principale del cancro ai polmoni nei non fumatori. I livelli di radon variano notevolmente da famiglia a famiglia.
L'EPA ha condotto uno studio sui livelli di radon in 80.000 case. Due importanti predittori sono: 1. Misurazione nel seminterrato o al primo piano (radon più alto negli scantinati) 2. Livello di uranio della contea (correlazione positiva con i livelli di radon)
Ci concentreremo sulla modellazione dei livelli di radon in Minnesota. La gerarchia in questo esempio è costituita dalle famiglie all'interno di ogni contea.
3 Gestione dei dati
In questa sezione si ottiene il radon
set di dati e fare un po 'di pre-elaborazione minima.
def load_and_preprocess_radon_dataset(state='MN'):
"""Preprocess Radon dataset as done in "Bayesian Data Analysis" book.
We filter to Minnesota data (919 examples) and preprocess to obtain the
following features:
- `log_uranium_ppm`: Log of soil uranium measurements.
- `county`: Name of county in which the measurement was taken.
- `floor`: Floor of house (0 for basement, 1 for first floor) on which the
measurement was taken.
The target variable is `log_radon`, the log of the Radon measurement in the
house.
"""
ds = tfds.load('radon', split='train')
radon_data = tfds.as_dataframe(ds)
radon_data.rename(lambda s: s[9:] if s.startswith('feat') else s, axis=1, inplace=True)
df = radon_data[radon_data.state==state.encode()].copy()
# For any missing or invalid activity readings, we'll use a value of `0.1`.
df['radon'] = df.activity.apply(lambda x: x if x > 0. else 0.1)
# Make county names look nice.
df['county'] = df.county.apply(lambda s: s.decode()).str.strip().str.title()
# Remap categories to start from 0 and end at max(category).
county_name = sorted(df.county.unique())
df['county'] = df.county.astype(
pd.api.types.CategoricalDtype(categories=county_name)).cat.codes
county_name = list(map(str.strip, county_name))
df['log_radon'] = df['radon'].apply(np.log)
df['log_uranium_ppm'] = df['Uppm'].apply(np.log)
df = df[['idnum', 'log_radon', 'floor', 'county', 'log_uranium_ppm']]
return df, county_name
radon, county_name = load_and_preprocess_radon_dataset()
num_counties = len(county_name)
num_observations = len(radon)
# Create copies of variables as Tensors.
county = tf.convert_to_tensor(radon['county'], dtype=tf.int32)
floor = tf.convert_to_tensor(radon['floor'], dtype=tf.float32)
log_radon = tf.convert_to_tensor(radon['log_radon'], dtype=tf.float32)
log_uranium = tf.convert_to_tensor(radon['log_uranium_ppm'], dtype=tf.float32)
radon.head()
Distribuzione dei livelli di radon (scala logaritmica):
plt.hist(log_radon.numpy(), bins=25, edgecolor='white')
plt.xlabel("Histogram of Radon levels (Log Scale)")
plt.show()
4 Approcci Convenzionali
Le due alternative convenzionali alla modellazione dell'esposizione al radon rappresentano i due estremi del compromesso bias-varianza:
Raggruppamento completo:
Tratta tutte le contee allo stesso modo e stima un singolo livello di radon.
\[y_i = \alpha + \beta x_i + \epsilon_i\]
Nessun raggruppamento:
Modella il radon in ogni contea in modo indipendente.
\(y_i = \alpha_{j[i]} + \beta x_i + \epsilon_i\) dove \(j = 1,\ldots,85\)
L'errori \(\epsilon_i\) può rappresentare errore di misura, temporali variazione all'interno-casa, o variazione tra le case.
4.1 Modello completo di pooling
pgm = daft.PGM([7, 3.5], node_unit=1.2)
pgm.add_node(
daft.Node(
"alpha_prior",
r"$\mathcal{N}(0, 10^5)$",
1,
3,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"beta_prior",
r"$\mathcal{N}(0, 10^5)$",
2.5,
3,
fixed=True,
offset=(10, 5)))
pgm.add_node(
daft.Node(
"sigma_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
4.5,
3,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("alpha", r"$\alpha$", 1, 2))
pgm.add_node(daft.Node("beta", r"$\beta$", 2.5, 2))
pgm.add_node(daft.Node("sigma", r"$\sigma$", 4.5, 2))
pgm.add_node(
daft.Node(
"y_i", r"$y_i \sim \mathcal{N}$", 3, 1, scale=1.25, observed=True))
pgm.add_edge("alpha_prior", "alpha")
pgm.add_edge("beta_prior", "beta")
pgm.add_edge("sigma_prior", "sigma")
pgm.add_edge("sigma", "y_i")
pgm.add_edge("alpha", "y_i")
pgm.add_edge("beta", "y_i")
pgm.add_plate(daft.Plate([2.3, 0.1, 1.4, 1.4], "$i = 1:919$"))
pgm.render()
plt.show()
Di seguito, adattiamo il modello di pooling completo utilizzando il Monte Carlo hamiltoniano.
@tf.function
def affine(x, kernel_diag, bias=tf.zeros([])):
"""`kernel_diag * x + bias` with broadcasting."""
kernel_diag = tf.ones_like(x) * kernel_diag
bias = tf.ones_like(x) * bias
return x * kernel_diag + bias
def pooled_model(floor):
"""Creates a joint distribution representing our generative process."""
return tfd.JointDistributionSequential([
tfd.Normal(loc=0., scale=1e5), # alpha
tfd.Normal(loc=0., scale=1e5), # beta
tfd.HalfCauchy(loc=0., scale=5), # sigma
lambda s, b1, b0: tfd.MultivariateNormalDiag( # y
loc=affine(floor, b1[..., tf.newaxis], b0[..., tf.newaxis]),
scale_identity_multiplier=s)
])
@tf.function
def pooled_log_prob(alpha, beta, sigma):
"""Computes `joint_log_prob` pinned at `log_radon`."""
return pooled_model(floor).log_prob([alpha, beta, sigma, log_radon])
@tf.function
def sample_pooled(num_chains, num_results, num_burnin_steps, num_observations):
"""Samples from the pooled model."""
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=pooled_log_prob,
num_leapfrog_steps=10,
step_size=0.005)
initial_state = [
tf.zeros([num_chains], name='init_alpha'),
tf.zeros([num_chains], name='init_beta'),
tf.ones([num_chains], name='init_sigma')
]
# Constrain `sigma` to the positive real axis. Other variables are
# unconstrained.
unconstraining_bijectors = [
tfb.Identity(), # alpha
tfb.Identity(), # beta
tfb.Exp() # sigma
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
PooledModel = collections.namedtuple('PooledModel', ['alpha', 'beta', 'sigma'])
samples, acceptance_probs = sample_pooled(
num_chains=4,
num_results=1000,
num_burnin_steps=1000,
num_observations=num_observations)
print('Acceptance Probabilities for each chain: ', acceptance_probs.numpy())
pooled_samples = PooledModel._make(samples)
Acceptance Probabilities for each chain: [0.997 0.99 0.997 0.995]
for var, var_samples in pooled_samples._asdict().items():
print('R-hat for ', var, ':\t',
tfp.mcmc.potential_scale_reduction(var_samples).numpy())
R-hat for alpha : 1.0046891 R-hat for beta : 1.0128309 R-hat for sigma : 1.0010641
def reduce_samples(var_samples, reduce_fn):
"""Reduces across leading two dims using reduce_fn."""
# Collapse the first two dimensions, typically (num_chains, num_samples), and
# compute np.mean or np.std along the remaining axis.
if isinstance(var_samples, tf.Tensor):
var_samples = var_samples.numpy() # convert to numpy array
var_samples = np.reshape(var_samples, (-1,) + var_samples.shape[2:])
return np.apply_along_axis(reduce_fn, axis=0, arr=var_samples)
sample_mean = lambda samples : reduce_samples(samples, np.mean)
Tracciare le stime puntuali della pendenza e dell'intercetta per il modello di pooling completo.
LinearEstimates = collections.namedtuple('LinearEstimates',
['intercept', 'slope'])
pooled_estimate = LinearEstimates(
intercept=sample_mean(pooled_samples.alpha),
slope=sample_mean(pooled_samples.beta)
)
plt.scatter(radon.floor, radon.log_radon)
xvals = np.linspace(-0.2, 1.2)
plt.ylabel('Radon level (Log Scale)')
plt.xticks([0, 1], ['Basement', 'First Floor'])
plt.plot(xvals, pooled_estimate.intercept + pooled_estimate.slope * xvals, 'r--')
plt.show()
Funzione di utilità per tracciare tracce di variabili campionate.
def plot_traces(var_name, samples, num_chains):
if isinstance(samples, tf.Tensor):
samples = samples.numpy() # convert to numpy array
fig, axes = plt.subplots(1, 2, figsize=(14, 1.5), sharex='col', sharey='col')
for chain in range(num_chains):
axes[0].plot(samples[:, chain], alpha=0.7)
axes[0].title.set_text("'{}' trace".format(var_name))
sns.kdeplot(samples[:, chain], ax=axes[1], shade=False)
axes[1].title.set_text("'{}' distribution".format(var_name))
axes[0].set_xlabel('Iteration')
axes[1].set_xlabel(var_name)
plt.show()
for var, var_samples in pooled_samples._asdict().items():
plot_traces(var, samples=var_samples, num_chains=4)
Successivamente, stimiamo i livelli di radon per ogni contea nel modello unpool.
4.2 Modello non in pool
pgm = daft.PGM([7, 3.5], node_unit=1.2)
pgm.add_node(
daft.Node(
"alpha_prior",
r"$\mathcal{N}(0, 10^5)$",
1,
3,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"beta_prior",
r"$\mathcal{N}(0, 10^5)$",
2.5,
3,
fixed=True,
offset=(10, 5)))
pgm.add_node(
daft.Node(
"sigma_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
4.5,
3,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("alpha", r"$\alpha$", 1, 2))
pgm.add_node(daft.Node("beta", r"$\beta$", 2.5, 2))
pgm.add_node(daft.Node("sigma", r"$\sigma$", 4.5, 2))
pgm.add_node(
daft.Node(
"y_i", r"$y_i \sim \mathcal{N}$", 3, 1, scale=1.25, observed=True))
pgm.add_edge("alpha_prior", "alpha")
pgm.add_edge("beta_prior", "beta")
pgm.add_edge("sigma_prior", "sigma")
pgm.add_edge("sigma", "y_i")
pgm.add_edge("alpha", "y_i")
pgm.add_edge("beta", "y_i")
pgm.add_plate(daft.Plate([0.3, 1.1, 1.4, 1.4], "$i = 1:85$"))
pgm.add_plate(daft.Plate([2.3, 0.1, 1.4, 1.4], "$i = 1:919$"))
pgm.render()
plt.show()
def unpooled_model(floor, county):
"""Creates a joint distribution for the unpooled model."""
return tfd.JointDistributionSequential([
tfd.MultivariateNormalDiag( # alpha
loc=tf.zeros([num_counties]), scale_identity_multiplier=1e5),
tfd.Normal(loc=0., scale=1e5), # beta
tfd.HalfCauchy(loc=0., scale=5), # sigma
lambda s, b1, b0: tfd.MultivariateNormalDiag( # y
loc=affine(
floor, b1[..., tf.newaxis], tf.gather(b0, county, axis=-1)),
scale_identity_multiplier=s)
])
@tf.function
def unpooled_log_prob(beta0, beta1, sigma):
"""Computes `joint_log_prob` pinned at `log_radon`."""
return (
unpooled_model(floor, county).log_prob([beta0, beta1, sigma, log_radon]))
@tf.function
def sample_unpooled(num_chains, num_results, num_burnin_steps):
"""Samples from the unpooled model."""
# Initialize the HMC transition kernel.
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=unpooled_log_prob,
num_leapfrog_steps=10,
step_size=0.025)
initial_state = [
tf.zeros([num_chains, num_counties], name='init_beta0'),
tf.zeros([num_chains], name='init_beta1'),
tf.ones([num_chains], name='init_sigma')
]
# Contrain `sigma` to the positive real axis. Other variables are
# unconstrained.
unconstraining_bijectors = [
tfb.Identity(), # alpha
tfb.Identity(), # beta
tfb.Exp() # sigma
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
UnpooledModel = collections.namedtuple('UnpooledModel',
['alpha', 'beta', 'sigma'])
samples, acceptance_probs = sample_unpooled(
num_chains=4, num_results=1000, num_burnin_steps=1000)
print('Acceptance Probabilities: ', acceptance_probs.numpy())
unpooled_samples = UnpooledModel._make(samples)
print('R-hat for beta:',
tfp.mcmc.potential_scale_reduction(unpooled_samples.beta).numpy())
print('R-hat for sigma:',
tfp.mcmc.potential_scale_reduction(unpooled_samples.sigma).numpy())
Acceptance Probabilities: [0.892 0.897 0.911 0.91 ] R-hat for beta: 1.0079623 R-hat for sigma: 1.0059084
plot_traces(var_name='beta', samples=unpooled_samples.beta, num_chains=4)
plot_traces(var_name='sigma', samples=unpooled_samples.sigma, num_chains=4)
Ecco i valori attesi della contea non raggruppati per l'intercetta insieme agli intervalli di credibilità del 95% per ciascuna catena. Riportiamo anche il valore R-hat per la stima di ciascuna contea.
Funzione di utilità per gli appezzamenti forestali.
def forest_plot(num_chains, num_vars, var_name, var_labels, samples):
fig, axes = plt.subplots(
1, 2, figsize=(12, 15), sharey=True, gridspec_kw={'width_ratios': [3, 1]})
for var_idx in range(num_vars):
values = samples[..., var_idx]
rhat = tfp.mcmc.diagnostic.potential_scale_reduction(values).numpy()
meds = np.median(values, axis=-2)
los = np.percentile(values, 5, axis=-2)
his = np.percentile(values, 95, axis=-2)
for i in range(num_chains):
height = 0.1 + 0.3 * var_idx + 0.05 * i
axes[0].plot([los[i], his[i]], [height, height], 'C0-', lw=2, alpha=0.5)
axes[0].plot([meds[i]], [height], 'C0o', ms=1.5)
axes[1].plot([rhat], [height], 'C0o', ms=4)
axes[0].set_yticks(np.linspace(0.2, 0.3, num_vars))
axes[0].set_ylim(0, 26)
axes[0].grid(which='both')
axes[0].invert_yaxis()
axes[0].set_yticklabels(var_labels)
axes[0].xaxis.set_label_position('top')
axes[0].set(xlabel='95% Credible Intervals for {}'.format(var_name))
axes[1].set_xticks([1, 2])
axes[1].set_xlim(0.95, 2.05)
axes[1].grid(which='both')
axes[1].set(xlabel='R-hat')
axes[1].xaxis.set_label_position('top')
plt.show()
forest_plot(
num_chains=4,
num_vars=num_counties,
var_name='alpha',
var_labels=county_name,
samples=unpooled_samples.alpha.numpy())
Possiamo tracciare le stime ordinate per identificare le contee con alti livelli di radon:
unpooled_intercepts = reduce_samples(unpooled_samples.alpha, np.mean)
unpooled_intercepts_se = reduce_samples(unpooled_samples.alpha, np.std)
def plot_ordered_estimates():
means = pd.Series(unpooled_intercepts, index=county_name)
std_errors = pd.Series(unpooled_intercepts_se, index=county_name)
order = means.sort_values().index
plt.plot(range(num_counties), means[order], '.')
for i, m, se in zip(range(num_counties), means[order], std_errors[order]):
plt.plot([i, i], [m - se, m + se], 'C0-')
plt.xlabel('Ordered county')
plt.ylabel('Radon estimate')
plt.show()
plot_ordered_estimates()
Funzione di utilità per tracciare le stime per un insieme campione di contee.
def plot_estimates(linear_estimates, labels, sample_counties):
fig, axes = plt.subplots(2, 4, figsize=(12, 6), sharey=True, sharex=True)
axes = axes.ravel()
intercepts_indexed = []
slopes_indexed = []
for intercepts, slopes in linear_estimates:
intercepts_indexed.append(pd.Series(intercepts, index=county_name))
slopes_indexed.append(pd.Series(slopes, index=county_name))
markers = ['-', 'r--', 'k:']
sample_county_codes = [county_name.index(c) for c in sample_counties]
for i, c in enumerate(sample_county_codes):
y = radon.log_radon[radon.county == c]
x = radon.floor[radon.county == c]
axes[i].scatter(
x + np.random.randn(len(x)) * 0.01, y, alpha=0.4, label='Log Radon')
# Plot both models and data
xvals = np.linspace(-0.2, 1.2)
for k in range(len(intercepts_indexed)):
axes[i].plot(
xvals,
intercepts_indexed[k][c] + slopes_indexed[k][c] * xvals,
markers[k],
label=labels[k])
axes[i].set_xticks([0, 1])
axes[i].set_xticklabels(['Basement', 'First Floor'])
axes[i].set_ylim(-1, 3)
axes[i].set_title(sample_counties[i])
if not i % 2:
axes[i].set_ylabel('Log Radon level')
axes[3].legend(bbox_to_anchor=(1.05, 0.9), borderaxespad=0.)
plt.show()
Di seguito sono riportati i confronti visivi tra le stime raggruppate e non raggruppate per un sottoinsieme di contee che rappresentano un intervallo di dimensioni del campione.
unpooled_estimates = LinearEstimates(
sample_mean(unpooled_samples.alpha),
sample_mean(unpooled_samples.beta)
)
sample_counties = ('Lac Qui Parle', 'Aitkin', 'Koochiching', 'Douglas', 'Clay',
'Stearns', 'Ramsey', 'St Louis')
plot_estimates(
linear_estimates=[unpooled_estimates, pooled_estimate],
labels=['Unpooled Estimates', 'Pooled Estimates'],
sample_counties=sample_counties)
Nessuno di questi modelli è soddisfacente:
- se stiamo cercando di identificare le contee ad alto contenuto di radon, il raggruppamento non è utile.
- non ci fidiamo delle stime unpool estreme prodotte da modelli che utilizzano poche osservazioni.
5 Modelli multilivello e gerarchici
Quando uniamo i nostri dati, perdiamo l'informazione che diversi punti dati provenivano da diverse contee. Questo significa che ogni radon
osservazione -level viene campionato dalla stessa distribuzione di probabilità. Tale modello non apprende alcuna variazione nell'unità di campionamento inerente a un gruppo (ad es. una contea). Tiene conto solo della varianza campionaria.
mpl.rc("font", size=18)
pgm = daft.PGM([13.6, 2.2], origin=[1.15, 1.0], node_ec="none")
pgm.add_node(daft.Node("parameter", r"parameter", 2.0, 3))
pgm.add_node(daft.Node("observations", r"observations", 2.0, 2))
pgm.add_node(daft.Node("theta", r"$\theta$", 5.5, 3))
pgm.add_node(daft.Node("y_0", r"$y_0$", 4, 2))
pgm.add_node(daft.Node("y_1", r"$y_1$", 5, 2))
pgm.add_node(daft.Node("dots", r"$\cdots$", 6, 2))
pgm.add_node(daft.Node("y_k", r"$y_k$", 7, 2))
pgm.add_edge("theta", "y_0")
pgm.add_edge("theta", "y_1")
pgm.add_edge("theta", "y_k")
pgm.render()
plt.show()
Quando analizziamo i dati non raggruppati, indichiamo che vengono campionati indipendentemente da modelli separati. All'estremo opposto rispetto al caso raggruppato, questo approccio afferma che le differenze tra le unità di campionamento sono troppo grandi per combinarle:
mpl.rc("font", size=18)
pgm = daft.PGM([13.6, 2.2], origin=[1.15, 1.0], node_ec="none")
pgm.add_node(daft.Node("parameter", r"parameter", 2.0, 3))
pgm.add_node(daft.Node("observations", r"observations", 2.0, 2))
pgm.add_node(daft.Node("theta_0", r"$\theta_0$", 4, 3))
pgm.add_node(daft.Node("theta_1", r"$\theta_1$", 5, 3))
pgm.add_node(daft.Node("theta_dots", r"$\cdots$", 6, 3))
pgm.add_node(daft.Node("theta_k", r"$\theta_k$", 7, 3))
pgm.add_node(daft.Node("y_0", r"$y_0$", 4, 2))
pgm.add_node(daft.Node("y_1", r"$y_1$", 5, 2))
pgm.add_node(daft.Node("y_dots", r"$\cdots$", 6, 2))
pgm.add_node(daft.Node("y_k", r"$y_k$", 7, 2))
pgm.add_edge("theta_0", "y_0")
pgm.add_edge("theta_1", "y_1")
pgm.add_edge("theta_k", "y_k")
pgm.render()
plt.show()
In un modello gerarchico, i parametri sono visti come un campione da una distribuzione di parametri della popolazione. Pertanto, li consideriamo né completamente diversi né esattamente uguali. Questo è noto come pool parziale.
mpl.rc("font", size=18)
pgm = daft.PGM([13.6, 3.4], origin=[1.15, 1.0], node_ec="none")
pgm.add_node(daft.Node("model", r"model", 2.0, 4))
pgm.add_node(daft.Node("parameter", r"parameter", 2.0, 3))
pgm.add_node(daft.Node("observations", r"observations", 2.0, 2))
pgm.add_node(daft.Node("mu_sigma", r"$\mu,\sigma^2$", 5.5, 4))
pgm.add_node(daft.Node("theta_0", r"$\theta_0$", 4, 3))
pgm.add_node(daft.Node("theta_1", r"$\theta_1$", 5, 3))
pgm.add_node(daft.Node("theta_dots", r"$\cdots$", 6, 3))
pgm.add_node(daft.Node("theta_k", r"$\theta_k$", 7, 3))
pgm.add_node(daft.Node("y_0", r"$y_0$", 4, 2))
pgm.add_node(daft.Node("y_1", r"$y_1$", 5, 2))
pgm.add_node(daft.Node("y_dots", r"$\cdots$", 6, 2))
pgm.add_node(daft.Node("y_k", r"$y_k$", 7, 2))
pgm.add_edge("mu_sigma", "theta_0")
pgm.add_edge("mu_sigma", "theta_1")
pgm.add_edge("mu_sigma", "theta_k")
pgm.add_edge("theta_0", "y_0")
pgm.add_edge("theta_1", "y_1")
pgm.add_edge("theta_k", "y_k")
pgm.render()
plt.show()
5.1 Raggruppamento parziale
Il modello di pooling parziale più semplice per il set di dati sul radon domestico è quello che stima semplicemente i livelli di radon, senza alcun predittore né a livello di gruppo né a livello individuale. Un esempio di predittore a livello individuale è se il punto dati proviene dal seminterrato o dal primo piano. Un predittore a livello di gruppo può essere il livello medio di uranio a livello di contea.
Un modello di raggruppamento parziale rappresenta un compromesso tra gli estremi raggruppati e non raggruppati, approssimativamente una media ponderata (basata sulla dimensione del campione) delle stime della contea non raggruppate e delle stime raggruppate.
Lasciate \(\hat{\alpha}_j\) essere il livello stimato di log-radon nella contea di \(j\). È solo un'intercettazione; ignoriamo le pendenze per ora. \(n_j\) è il numero di osservazioni contea \(j\). \(\sigma_{\alpha}\) e \(\sigma_y\) sono varianza all'interno del parametro e la varianza di campionamento rispettivamente. Quindi un modello di pooling parziale potrebbe postulare:
\[\hat{\alpha}_j \approx \frac{(n_j/\sigma_y^2)\bar{y}_j + (1/\sigma_{\alpha}^2)\bar{y} }{(n_j/\sigma_y^2) + (1/\sigma_{\alpha}^2)}\]
Ci aspettiamo quanto segue quando si utilizza il pooling parziale:
- Le stime per le contee con campioni di dimensioni inferiori si ridurranno verso la media statale.
- Le stime per le contee con campioni di dimensioni maggiori saranno più vicine alle stime per le contee non raggruppate.
mpl.rc("font", size=12)
pgm = daft.PGM([7, 4.5], node_unit=1.2)
pgm.add_node(
daft.Node(
"mu_a_prior",
r"$\mathcal{N}(0, 10^5)$",
1,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"sigma_a_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
3,
4,
fixed=True,
offset=(10, 5)))
pgm.add_node(
daft.Node(
"sigma_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
4,
3,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("mu_a", r"$\mu_a$", 1, 3))
pgm.add_node(daft.Node("sigma_a", r"$\sigma_a$", 3, 3))
pgm.add_node(daft.Node("a", r"$a \sim \mathcal{N}$", 2, 2, scale=1.25))
pgm.add_node(daft.Node("sigma_y", r"$\sigma_y$", 4, 2))
pgm.add_node(
daft.Node(
"y_i", r"$y_i \sim \mathcal{N}$", 3.25, 1, scale=1.25, observed=True))
pgm.add_edge("mu_a_prior", "mu_a")
pgm.add_edge("sigma_a_prior", "sigma_a")
pgm.add_edge("mu_a", "a")
pgm.add_edge("sigma_a", "a")
pgm.add_edge("sigma_prior", "sigma_y")
pgm.add_edge("sigma_y", "y_i")
pgm.add_edge("a", "y_i")
pgm.add_plate(daft.Plate([1.4, 1.2, 1.2, 1.4], "$i = 1:85$"))
pgm.add_plate(daft.Plate([2.65, 0.2, 1.2, 1.4], "$i = 1:919$"))
pgm.render()
plt.show()
def partial_pooling_model(county):
"""Creates a joint distribution for the partial pooling model."""
return tfd.JointDistributionSequential([
tfd.Normal(loc=0., scale=1e5), # mu_a
tfd.HalfCauchy(loc=0., scale=5), # sigma_a
lambda sigma_a, mu_a: tfd.MultivariateNormalDiag( # a
loc=mu_a[..., tf.newaxis] * tf.ones([num_counties])[tf.newaxis, ...],
scale_identity_multiplier=sigma_a),
tfd.HalfCauchy(loc=0., scale=5), # sigma_y
lambda sigma_y, a: tfd.MultivariateNormalDiag( # y
loc=tf.gather(a, county, axis=-1),
scale_identity_multiplier=sigma_y)
])
@tf.function
def partial_pooling_log_prob(mu_a, sigma_a, a, sigma_y):
"""Computes joint log prob pinned at `log_radon`."""
return partial_pooling_model(county).log_prob(
[mu_a, sigma_a, a, sigma_y, log_radon])
@tf.function
def sample_partial_pooling(num_chains, num_results, num_burnin_steps):
"""Samples from the partial pooling model."""
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=partial_pooling_log_prob,
num_leapfrog_steps=10,
step_size=0.01)
initial_state = [
tf.zeros([num_chains], name='init_mu_a'),
tf.ones([num_chains], name='init_sigma_a'),
tf.zeros([num_chains, num_counties], name='init_a'),
tf.ones([num_chains], name='init_sigma_y')
]
unconstraining_bijectors = [
tfb.Identity(), # mu_a
tfb.Exp(), # sigma_a
tfb.Identity(), # a
tfb.Exp() # sigma_y
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
PartialPoolingModel = collections.namedtuple(
'PartialPoolingModel', ['mu_a', 'sigma_a', 'a', 'sigma_y'])
samples, acceptance_probs = sample_partial_pooling(
num_chains=4, num_results=1000, num_burnin_steps=1000)
print('Acceptance Probabilities: ', acceptance_probs.numpy())
partial_pooling_samples = PartialPoolingModel._make(samples)
Acceptance Probabilities: [0.989 0.977 0.988 0.985]
for var in ['mu_a', 'sigma_a', 'sigma_y']:
print(
'R-hat for ', var, '\t:',
tfp.mcmc.potential_scale_reduction(getattr(partial_pooling_samples,
var)).numpy())
R-hat for mu_a : 1.0216417 R-hat for sigma_a : 1.0224565 R-hat for sigma_y : 1.0016255
partial_pooling_intercepts = reduce_samples(
partial_pooling_samples.a.numpy(), np.mean)
partial_pooling_intercepts_se = reduce_samples(
partial_pooling_samples.a.numpy(), np.std)
def plot_unpooled_vs_partial_pooling_estimates():
fig, axes = plt.subplots(1, 2, figsize=(14, 6), sharex=True, sharey=True)
# Order counties by number of observations (and add some jitter).
num_obs_per_county = (
radon.groupby('county')['idnum'].count().values.astype(np.float32))
num_obs_per_county += np.random.normal(scale=0.5, size=num_counties)
intercepts_list = [unpooled_intercepts, partial_pooling_intercepts]
intercepts_se_list = [unpooled_intercepts_se, partial_pooling_intercepts_se]
for ax, means, std_errors in zip(axes, intercepts_list, intercepts_se_list):
ax.plot(num_obs_per_county, means, 'C0.')
for n, m, se in zip(num_obs_per_county, means, std_errors):
ax.plot([n, n], [m - se, m + se], 'C1-', alpha=.5)
for ax in axes:
ax.set_xscale('log')
ax.set_xlabel('No. of Observations Per County')
ax.set_xlim(1, 100)
ax.set_ylabel('Log Radon Estimate (with Standard Error)')
ax.set_ylim(0, 3)
ax.hlines(partial_pooling_intercepts.mean(), .9, 125, 'k', '--', alpha=.5)
axes[0].set_title('Unpooled Estimates')
axes[1].set_title('Partially Pooled Estimates')
plot_unpooled_vs_partial_pooling_estimates()
Notare la differenza tra le stime non raggruppate e quelle parzialmente raggruppate, in particolare per campioni di dimensioni inferiori. I primi sono sia più estremi che più imprecisi.
5.2 Intercettazioni variabili
Consideriamo ora un modello più complesso che consente alle intercettazioni di variare da una contea all'altra, secondo un effetto casuale.
\(y_i = \alpha_{j[i]} + \beta x_{i} + \epsilon_i\) dove\(\epsilon_i \sim N(0, \sigma_y^2)\) e l'effetto casuale intercettazione:
\[\alpha_{j[i]} \sim N(\mu_{\alpha}, \sigma_{\alpha}^2)\]
La pendenza \(\beta\), che permette l'osservazione variare in funzione della posizione di misurazione (seminterrato o primo piano), è ancora un effetto fisso condivisa tra diverse contee.
Come con il modello unpooling, abbiamo impostato intercetta separato per ogni regione, ma piuttosto che adattare minimi quadrati modelli di regressione separati per ogni regione, la forza multilivello azioni modellazione tra contee, consentendo un più ragionevole inferenza in paesi con pochi dati.
mpl.rc("font", size=12)
pgm = daft.PGM([7, 4.5], node_unit=1.2)
pgm.add_node(
daft.Node(
"mu_a_prior",
r"$\mathcal{N}(0, 10^5)$",
1,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"sigma_a_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
3,
4,
fixed=True,
offset=(10, 5)))
pgm.add_node(
daft.Node(
"b_prior",
r"$\mathcal{N}(0, 10^5)$",
4,
3.5,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("b", r"$b$", 4, 2.5))
pgm.add_node(
daft.Node(
"sigma_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
6,
3.5,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("mu_a", r"$\mu_a$", 1, 3))
pgm.add_node(daft.Node("sigma_a", r"$\sigma_a$", 3, 3))
pgm.add_node(daft.Node("a", r"$a \sim \mathcal{N}$", 2, 2, scale=1.25))
pgm.add_node(daft.Node("sigma_y", r"$\sigma_y$", 6, 2.5))
pgm.add_node(
daft.Node(
"y_i", r"$y_i \sim \mathcal{N}$", 4, 1, scale=1.25, observed=True))
pgm.add_edge("mu_a_prior", "mu_a")
pgm.add_edge("sigma_a_prior", "sigma_a")
pgm.add_edge("mu_a", "a")
pgm.add_edge("b_prior", "b")
pgm.add_edge("sigma_a", "a")
pgm.add_edge("sigma_prior", "sigma_y")
pgm.add_edge("sigma_y", "y_i")
pgm.add_edge("a", "y_i")
pgm.add_edge("b", "y_i")
pgm.add_plate(daft.Plate([1.4, 1.2, 1.2, 1.4], "$i = 1:85$"))
pgm.add_plate(daft.Plate([3.4, 0.2, 1.2, 1.4], "$i = 1:919$"))
pgm.render()
plt.show()
def varying_intercept_model(floor, county):
"""Creates a joint distribution for the varying intercept model."""
return tfd.JointDistributionSequential([
tfd.Normal(loc=0., scale=1e5), # mu_a
tfd.HalfCauchy(loc=0., scale=5), # sigma_a
lambda sigma_a, mu_a: tfd.MultivariateNormalDiag( # a
loc=affine(tf.ones([num_counties]), mu_a[..., tf.newaxis]),
scale_identity_multiplier=sigma_a),
tfd.Normal(loc=0., scale=1e5), # b
tfd.HalfCauchy(loc=0., scale=5), # sigma_y
lambda sigma_y, b, a: tfd.MultivariateNormalDiag( # y
loc=affine(floor, b[..., tf.newaxis], tf.gather(a, county, axis=-1)),
scale_identity_multiplier=sigma_y)
])
def varying_intercept_log_prob(mu_a, sigma_a, a, b, sigma_y):
"""Computes joint log prob pinned at `log_radon`."""
return varying_intercept_model(floor, county).log_prob(
[mu_a, sigma_a, a, b, sigma_y, log_radon])
@tf.function
def sample_varying_intercepts(num_chains, num_results, num_burnin_steps):
"""Samples from the varying intercepts model."""
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=varying_intercept_log_prob,
num_leapfrog_steps=10,
step_size=0.01)
initial_state = [
tf.zeros([num_chains], name='init_mu_a'),
tf.ones([num_chains], name='init_sigma_a'),
tf.zeros([num_chains, num_counties], name='init_a'),
tf.zeros([num_chains], name='init_b'),
tf.ones([num_chains], name='init_sigma_y')
]
unconstraining_bijectors = [
tfb.Identity(), # mu_a
tfb.Exp(), # sigma_a
tfb.Identity(), # a
tfb.Identity(), # b
tfb.Exp() # sigma_y
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
VaryingInterceptsModel = collections.namedtuple(
'VaryingInterceptsModel', ['mu_a', 'sigma_a', 'a', 'b', 'sigma_y'])
samples, acceptance_probs = sample_varying_intercepts(
num_chains=4, num_results=1000, num_burnin_steps=1000)
print('Acceptance Probabilities: ', acceptance_probs.numpy())
varying_intercepts_samples = VaryingInterceptsModel._make(samples)
Acceptance Probabilities: [0.978 0.987 0.982 0.984]
for var in ['mu_a', 'sigma_a', 'b', 'sigma_y']:
print(
'R-hat for ', var, ': ',
tfp.mcmc.potential_scale_reduction(
getattr(varying_intercepts_samples, var)).numpy())
R-hat for mu_a : 1.1099764 R-hat for sigma_a : 1.1058794 R-hat for b : 1.0448593 R-hat for sigma_y : 1.0019052
varying_intercepts_estimates = LinearEstimates(
sample_mean(varying_intercepts_samples.a),
sample_mean(varying_intercepts_samples.b))
sample_counties = ('Lac Qui Parle', 'Aitkin', 'Koochiching', 'Douglas', 'Clay',
'Stearns', 'Ramsey', 'St Louis')
plot_estimates(
linear_estimates=[
unpooled_estimates, pooled_estimate, varying_intercepts_estimates
],
labels=['Unpooled', 'Pooled', 'Varying Intercepts'],
sample_counties=sample_counties)
def plot_posterior(var_name, var_samples):
if isinstance(var_samples, tf.Tensor):
var_samples = var_samples.numpy() # convert to numpy array
fig = plt.figure(figsize=(10, 3))
ax = fig.add_subplot(111)
ax.hist(var_samples.flatten(), bins=40, edgecolor='white')
sample_mean = var_samples.mean()
ax.text(
sample_mean,
100,
'mean={:.3f}'.format(sample_mean),
color='white',
fontsize=12)
ax.set_xlabel('posterior of ' + var_name)
plt.show()
plot_posterior('b', varying_intercepts_samples.b)
plot_posterior('sigma_a', varying_intercepts_samples.sigma_a)
La stima del coefficiente pavimento è di circa -0.69, che può essere interpretato come case senza scantinati aventi circa la metà (\(\exp(-0.69) = 0.50\)) i livelli di radon di quelli con scantinati, dopo la contabilizzazione per contea.
for var in ['b']:
var_samples = getattr(varying_intercepts_samples, var)
mean = var_samples.numpy().mean()
std = var_samples.numpy().std()
r_hat = tfp.mcmc.potential_scale_reduction(var_samples).numpy()
n_eff = tfp.mcmc.effective_sample_size(var_samples).numpy().sum()
print('var: ', var, ' mean: ', mean, ' std: ', std, ' n_eff: ', n_eff,
' r_hat: ', r_hat)
var: b mean: -0.6972574 std: 0.06966117 n_eff: 397.94327 r_hat: 1.0448593
def plot_intercepts_and_slopes(linear_estimates, title):
xvals = np.arange(2)
intercepts = np.ones([num_counties]) * linear_estimates.intercept
slopes = np.ones([num_counties]) * linear_estimates.slope
fig, ax = plt.subplots()
for c in range(num_counties):
ax.plot(xvals, intercepts[c] + slopes[c] * xvals, 'bo-', alpha=0.4)
plt.xlim(-0.2, 1.2)
ax.set_xticks([0, 1])
ax.set_xticklabels(['Basement', 'First Floor'])
ax.set_ylabel('Log Radon level')
plt.title(title)
plt.show()
plot_intercepts_and_slopes(varying_intercepts_estimates,
'Log Radon Estimates (Varying Intercepts)')
5.3 Pendenze variabili
In alternativa, possiamo ipotizzare un modello che permetta alle contee di variare a seconda di come il luogo di misurazione (seminterrato o primo piano) influenzi la lettura del radon. In questo caso l'intercetta \(\alpha\) è condivisa tra contee.
\[y_i = \alpha + \beta_{j[i]} x_{i} + \epsilon_i\]
mpl.rc("font", size=12)
pgm = daft.PGM([10, 4.5], node_unit=1.2)
pgm.add_node(
daft.Node(
"mu_b_prior",
r"$\mathcal{N}(0, 10^5)$",
3.2,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"a_prior", r"$\mathcal{N}(0, 10^5)$", 2, 3, fixed=True, offset=(20, 5)))
pgm.add_node(daft.Node("a", r"$a$", 2, 2))
pgm.add_node(
daft.Node(
"sigma_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
4,
3.5,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("sigma_y", r"$\sigma_y$", 4, 2.5))
pgm.add_node(
daft.Node(
"mu_b_prior",
r"$\mathcal{N}(0, 10^5)$",
5,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"sigma_b_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
7,
4,
fixed=True,
offset=(10, 5)))
pgm.add_node(daft.Node("mu_b", r"$\mu_b$", 5, 3))
pgm.add_node(daft.Node("sigma_b", r"$\sigma_b$", 7, 3))
pgm.add_node(daft.Node("b", r"$b \sim \mathcal{N}$", 6, 2, scale=1.25))
pgm.add_node(
daft.Node(
"y_i", r"$y_i \sim \mathcal{N}$", 4, 1, scale=1.25, observed=True))
pgm.add_edge("a_prior", "a")
pgm.add_edge("mu_b_prior", "mu_b")
pgm.add_edge("sigma_b_prior", "sigma_b")
pgm.add_edge("mu_b", "b")
pgm.add_edge("sigma_b", "b")
pgm.add_edge("sigma_prior", "sigma_y")
pgm.add_edge("sigma_y", "y_i")
pgm.add_edge("a", "y_i")
pgm.add_edge("b", "y_i")
pgm.add_plate(daft.Plate([5.4, 1.2, 1.2, 1.4], "$i = 1:85$"))
pgm.add_plate(daft.Plate([3.4, 0.2, 1.2, 1.4], "$i = 1:919$"))
pgm.render()
plt.show()
def varying_slopes_model(floor, county):
"""Creates a joint distribution for the varying slopes model."""
return tfd.JointDistributionSequential([
tfd.Normal(loc=0., scale=1e5), # mu_b
tfd.HalfCauchy(loc=0., scale=5), # sigma_b
tfd.Normal(loc=0., scale=1e5), # a
lambda _, sigma_b, mu_b: tfd.MultivariateNormalDiag( # b
loc=affine(tf.ones([num_counties]), mu_b[..., tf.newaxis]),
scale_identity_multiplier=sigma_b),
tfd.HalfCauchy(loc=0., scale=5), # sigma_y
lambda sigma_y, b, a: tfd.MultivariateNormalDiag( # y
loc=affine(floor, tf.gather(b, county, axis=-1), a[..., tf.newaxis]),
scale_identity_multiplier=sigma_y)
])
def varying_slopes_log_prob(mu_b, sigma_b, a, b, sigma_y):
return varying_slopes_model(floor, county).log_prob(
[mu_b, sigma_b, a, b, sigma_y, log_radon])
@tf.function
def sample_varying_slopes(num_chains, num_results, num_burnin_steps):
"""Samples from the varying slopes model."""
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=varying_slopes_log_prob,
num_leapfrog_steps=25,
step_size=0.01)
initial_state = [
tf.zeros([num_chains], name='init_mu_b'),
tf.ones([num_chains], name='init_sigma_b'),
tf.zeros([num_chains], name='init_a'),
tf.zeros([num_chains, num_counties], name='init_b'),
tf.ones([num_chains], name='init_sigma_y')
]
unconstraining_bijectors = [
tfb.Identity(), # mu_b
tfb.Exp(), # sigma_b
tfb.Identity(), # a
tfb.Identity(), # b
tfb.Exp() # sigma_y
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
VaryingSlopesModel = collections.namedtuple(
'VaryingSlopesModel', ['mu_b', 'sigma_b', 'a', 'b', 'sigma_y'])
samples, acceptance_probs = sample_varying_slopes(
num_chains=4, num_results=1000, num_burnin_steps=1000)
print('Acceptance Probabilities: ', acceptance_probs.numpy())
varying_slopes_samples = VaryingSlopesModel._make(samples)
Acceptance Probabilities: [0.979 0.984 0.977 0.984]
for var in ['mu_b', 'sigma_b', 'a', 'sigma_y']:
print(
'R-hat for ', var, '\t: ',
tfp.mcmc.potential_scale_reduction(getattr(varying_slopes_samples,
var)).numpy())
R-hat for mu_b : 1.0770341 R-hat for sigma_b : 1.0634488 R-hat for a : 1.0133665 R-hat for sigma_y : 1.0011941
varying_slopes_estimates = LinearEstimates(
sample_mean(varying_slopes_samples.a),
sample_mean(varying_slopes_samples.b))
plot_intercepts_and_slopes(varying_slopes_estimates,
'Log Radon Estimates (Varying Slopes)')
5.4 Variazione delle intercettazioni e delle pendenze
Il modello più generale consente sia all'intercetta che alla pendenza di variare in base alla contea:
\[y_i = \alpha_{j[i]} + \beta_{j[i]} x_{i} + \epsilon_i\]
mpl.rc("font", size=12)
pgm = daft.PGM([10, 4.5], node_unit=1.2)
pgm.add_node(
daft.Node(
"mu_a_prior",
r"$\mathcal{N}(0, 10^5)$",
1,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"sigma_a_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
3,
4,
fixed=True,
offset=(10, 5)))
pgm.add_node(daft.Node("mu_a", r"$\mu_a$", 1, 3))
pgm.add_node(daft.Node("sigma_a", r"$\sigma_a$", 3, 3))
pgm.add_node(daft.Node("a", r"$a \sim \mathcal{N}$", 2, 2, scale=1.25))
pgm.add_node(
daft.Node(
"sigma_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
4,
3.5,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("sigma_y", r"$\sigma_y$", 4, 2.5))
pgm.add_node(
daft.Node(
"mu_b_prior",
r"$\mathcal{N}(0, 10^5)$",
5,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"sigma_b_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
7,
4,
fixed=True,
offset=(10, 5)))
pgm.add_node(daft.Node("mu_b", r"$\mu_b$", 5, 3))
pgm.add_node(daft.Node("sigma_b", r"$\sigma_b$", 7, 3))
pgm.add_node(daft.Node("b", r"$b \sim \mathcal{N}$", 6, 2, scale=1.25))
pgm.add_node(
daft.Node(
"y_i", r"$y_i \sim \mathcal{N}$", 4, 1, scale=1.25, observed=True))
pgm.add_edge("mu_a_prior", "mu_a")
pgm.add_edge("sigma_a_prior", "sigma_a")
pgm.add_edge("mu_a", "a")
pgm.add_edge("sigma_a", "a")
pgm.add_edge("mu_b_prior", "mu_b")
pgm.add_edge("sigma_b_prior", "sigma_b")
pgm.add_edge("mu_b", "b")
pgm.add_edge("sigma_b", "b")
pgm.add_edge("sigma_prior", "sigma_y")
pgm.add_edge("sigma_y", "y_i")
pgm.add_edge("a", "y_i")
pgm.add_edge("b", "y_i")
pgm.add_plate(daft.Plate([1.4, 1.2, 1.2, 1.4], "$i = 1:85$"))
pgm.add_plate(daft.Plate([5.4, 1.2, 1.2, 1.4], "$i = 1:85$"))
pgm.add_plate(daft.Plate([3.4, 0.2, 1.2, 1.4], "$i = 1:919$"))
pgm.render()
plt.show()
def varying_intercepts_and_slopes_model(floor, county):
"""Creates a joint distribution for the varying slope model."""
return tfd.JointDistributionSequential([
tfd.Normal(loc=0., scale=1e5), # mu_a
tfd.HalfCauchy(loc=0., scale=5), # sigma_a
tfd.Normal(loc=0., scale=1e5), # mu_b
tfd.HalfCauchy(loc=0., scale=5), # sigma_b
lambda sigma_b, mu_b, sigma_a, mu_a: tfd.MultivariateNormalDiag( # a
loc=affine(tf.ones([num_counties]), mu_a[..., tf.newaxis]),
scale_identity_multiplier=sigma_a),
lambda _, sigma_b, mu_b: tfd.MultivariateNormalDiag( # b
loc=affine(tf.ones([num_counties]), mu_b[..., tf.newaxis]),
scale_identity_multiplier=sigma_b),
tfd.HalfCauchy(loc=0., scale=5), # sigma_y
lambda sigma_y, b, a: tfd.MultivariateNormalDiag( # y
loc=affine(floor, tf.gather(b, county, axis=-1),
tf.gather(a, county, axis=-1)),
scale_identity_multiplier=sigma_y)
])
@tf.function
def varying_intercepts_and_slopes_log_prob(mu_a, sigma_a, mu_b, sigma_b, a, b,
sigma_y):
"""Computes joint log prob pinned at `log_radon`."""
return varying_intercepts_and_slopes_model(floor, county).log_prob(
[mu_a, sigma_a, mu_b, sigma_b, a, b, sigma_y, log_radon])
@tf.function
def sample_varying_intercepts_and_slopes(num_chains, num_results,
num_burnin_steps):
"""Samples from the varying intercepts and slopes model."""
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=varying_intercepts_and_slopes_log_prob,
num_leapfrog_steps=50,
step_size=0.01)
initial_state = [
tf.zeros([num_chains], name='init_mu_a'),
tf.ones([num_chains], name='init_sigma_a'),
tf.zeros([num_chains], name='init_mu_b'),
tf.ones([num_chains], name='init_sigma_b'),
tf.zeros([num_chains, num_counties], name='init_a'),
tf.zeros([num_chains, num_counties], name='init_b'),
tf.ones([num_chains], name='init_sigma_y')
]
unconstraining_bijectors = [
tfb.Identity(), # mu_a
tfb.Exp(), # sigma_a
tfb.Identity(), # mu_b
tfb.Exp(), # sigma_b
tfb.Identity(), # a
tfb.Identity(), # b
tfb.Exp() # sigma_y
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
VaryingInterceptsAndSlopesModel = collections.namedtuple(
'VaryingInterceptsAndSlopesModel',
['mu_a', 'sigma_a', 'mu_b', 'sigma_b', 'a', 'b', 'sigma_y'])
samples, acceptance_probs = sample_varying_intercepts_and_slopes(
num_chains=4, num_results=1000, num_burnin_steps=500)
print('Acceptance Probabilities: ', acceptance_probs.numpy())
varying_intercepts_and_slopes_samples = VaryingInterceptsAndSlopesModel._make(
samples)
Acceptance Probabilities: [0.988 0.985 0.992 0.938]
for var in ['mu_a', 'sigma_a', 'mu_b', 'sigma_b']:
print(
'R-hat for ', var, '\t: ',
tfp.mcmc.potential_scale_reduction(
getattr(varying_intercepts_and_slopes_samples, var)).numpy())
R-hat for mu_a : 1.010764 R-hat for sigma_a : 1.0078123 R-hat for mu_b : 1.0279609 R-hat for sigma_b : 1.3165458
varying_intercepts_and_slopes_estimates = LinearEstimates(
sample_mean(varying_intercepts_and_slopes_samples.a),
sample_mean(varying_intercepts_and_slopes_samples.b))
plot_intercepts_and_slopes(
varying_intercepts_and_slopes_estimates,
'Log Radon Estimates (Varying Intercepts and Slopes)')
forest_plot(
num_chains=4,
num_vars=num_counties,
var_name='a',
var_labels=county_name,
samples=varying_intercepts_and_slopes_samples.a.numpy())
forest_plot(
num_chains=4,
num_vars=num_counties,
var_name='b',
var_labels=county_name,
samples=varying_intercepts_and_slopes_samples.b.numpy())
6 Aggiunta di predittori a livello di gruppo
Un punto di forza principale dei modelli multilivello è la capacità di gestire i predittori su più livelli contemporaneamente. Se consideriamo il modello delle intercettazioni variabili sopra:
\(y_i = \alpha_{j[i]} + \beta x_{i} + \epsilon_i\) ci, invece di una semplice effetto casuale per descrivere variazione del valore atteso radon, specificare un altro modello di regressione con covariate a livello di contea. Qui, usiamo l'uranio contea lettura \(u_j\), che è pensato per essere collegato con i livelli di radon:
\(\alpha_j = \gamma_0 + \gamma_1 u_j + \zeta_j\)\(\zeta_j \sim N(0, \sigma_{\alpha}^2)\) Così, ora stiamo incorporando un predittore di livello casa (piano o in cantina) così come un predittore a livello di contea (uranio).
Si noti che il modello ha entrambe le variabili dell'indicatore per ogni contea, più una covariata a livello di contea. Nella regressione classica, ciò risulterebbe in collinearità. In un modello multilivello, il raggruppamento parziale delle intercettazioni verso il valore atteso del modello lineare a livello di gruppo lo evita.
Predittori a livello di gruppo servono anche a ridurre le variazioni a livello di gruppo\(\sigma_{\alpha}\). Un'implicazione importante di ciò è che la stima a livello di gruppo induce una maggiore aggregazione.
6.1 Modello gerarchico delle intercettazioni
mpl.rc("font", size=12)
pgm = daft.PGM([10, 4.5], node_unit=1.2)
pgm.add_node(
daft.Node(
"gamma_0_prior",
r"$\mathcal{N}(0, 10^5)$",
0.5,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(
daft.Node(
"gamma_1_prior",
r"$\mathcal{N}(0, 10^5)$",
1.5,
4,
fixed=True,
offset=(10, 5)))
pgm.add_node(daft.Node("gamma_0", r"$\gamma_0$", 0.5, 3))
pgm.add_node(daft.Node("gamma_1", r"$\gamma_1$", 1.5, 3))
pgm.add_node(
daft.Node(
"sigma_a_prior",
r"$\mathcal{N}(0, 10^5)$",
3,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(daft.Node("sigma_a", r"$\sigma_a$", 3, 3.5))
pgm.add_node(daft.Node("eps_a", r"$eps_a$", 3, 2.5, scale=1.25))
pgm.add_node(daft.Node("a", r"$a \sim \mathcal{Det}$", 1.5, 1.2, scale=1.5))
pgm.add_node(
daft.Node(
"sigma_prior",
r"$\mathrm{U}(0, 100)$",
4,
3.5,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("sigma_y", r"$\sigma_y$", 4, 2.5))
pgm.add_node(daft.Node("b_prior", r"$\mathcal{N}(0, 10^5)$", 5, 3, fixed=True))
pgm.add_node(daft.Node("b", r"$b$", 5, 2))
pgm.add_node(
daft.Node(
"y_i", r"$y_i \sim \mathcal{N}$", 4, 1, scale=1.25, observed=True))
pgm.add_edge("gamma_0_prior", "gamma_0")
pgm.add_edge("gamma_1_prior", "gamma_1")
pgm.add_edge("sigma_a_prior", "sigma_a")
pgm.add_edge("sigma_a", "eps_a")
pgm.add_edge("gamma_0", "a")
pgm.add_edge("gamma_1", "a")
pgm.add_edge("eps_a", "a")
pgm.add_edge("b_prior", "b")
pgm.add_edge("sigma_prior", "sigma_y")
pgm.add_edge("sigma_y", "y_i")
pgm.add_edge("a", "y_i")
pgm.add_edge("b", "y_i")
pgm.add_plate(daft.Plate([2.4, 1.7, 1.2, 1.4], "$i = 1:85$"))
pgm.add_plate(daft.Plate([0.9, 0.4, 1.2, 1.4], "$i = 1:919$"))
pgm.add_plate(daft.Plate([3.4, 0.2, 1.2, 1.4], "$i = 1:919$"))
pgm.render()
plt.show()
def hierarchical_intercepts_model(floor, county, log_uranium):
"""Creates a joint distribution for the varying slope model."""
return tfd.JointDistributionSequential([
tfd.HalfCauchy(loc=0., scale=5), # sigma_a
lambda sigma_a: tfd.MultivariateNormalDiag( # eps_a
loc=tf.zeros([num_counties]),
scale_identity_multiplier=sigma_a),
tfd.Normal(loc=0., scale=1e5), # gamma_0
tfd.Normal(loc=0., scale=1e5), # gamma_1
tfd.Normal(loc=0., scale=1e5), # b
tfd.Uniform(low=0., high=100), # sigma_y
lambda sigma_y, b, gamma_1, gamma_0, eps_a: tfd.
MultivariateNormalDiag( # y
loc=affine(
floor, b[..., tf.newaxis],
affine(log_uranium, gamma_1[..., tf.newaxis],
gamma_0[..., tf.newaxis]) + tf.gather(eps_a, county, axis=-1)),
scale_identity_multiplier=sigma_y)
])
def hierarchical_intercepts_log_prob(sigma_a, eps_a, gamma_0, gamma_1, b,
sigma_y):
"""Computes joint log prob pinned at `log_radon`."""
return hierarchical_intercepts_model(floor, county, log_uranium).log_prob(
[sigma_a, eps_a, gamma_0, gamma_1, b, sigma_y, log_radon])
@tf.function
def sample_hierarchical_intercepts(num_chains, num_results, num_burnin_steps):
"""Samples from the hierarchical intercepts model."""
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=hierarchical_intercepts_log_prob,
num_leapfrog_steps=10,
step_size=0.01)
initial_state = [
tf.ones([num_chains], name='init_sigma_a'),
tf.zeros([num_chains, num_counties], name='eps_a'),
tf.zeros([num_chains], name='init_gamma_0'),
tf.zeros([num_chains], name='init_gamma_1'),
tf.zeros([num_chains], name='init_b'),
tf.ones([num_chains], name='init_sigma_y')
]
unconstraining_bijectors = [
tfb.Exp(), # sigma_a
tfb.Identity(), # eps_a
tfb.Identity(), # gamma_0
tfb.Identity(), # gamma_0
tfb.Identity(), # b
# Maps reals to [0, 100].
tfb.Chain([tfb.Shift(shift=50.),
tfb.Scale(scale=50.),
tfb.Tanh()]) # sigma_y
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
HierarchicalInterceptsModel = collections.namedtuple(
'HierarchicalInterceptsModel',
['sigma_a', 'eps_a', 'gamma_0', 'gamma_1', 'b', 'sigma_y'])
samples, acceptance_probs = sample_hierarchical_intercepts(
num_chains=4, num_results=2000, num_burnin_steps=500)
print('Acceptance Probabilities: ', acceptance_probs.numpy())
hierarchical_intercepts_samples = HierarchicalInterceptsModel._make(samples)
Acceptance Probabilities: [0.9615 0.941 0.955 0.95 ]
for var in ['sigma_a', 'gamma_0', 'gamma_1', 'b', 'sigma_y']:
print(
'R-hat for', var, ':',
tfp.mcmc.potential_scale_reduction(
getattr(hierarchical_intercepts_samples, var)).numpy())
R-hat for sigma_a : 1.0469627 R-hat for gamma_0 : 1.0016835 R-hat for gamma_1 : 1.0097923 R-hat for b : 1.0014259 R-hat for sigma_y : 1.0025403
def plot_hierarchical_intercepts():
mean_and_var = lambda x : [reduce_samples(x, fn) for fn in [np.mean, np.var]]
gamma_0_mean, gamma_0_var = mean_and_var(
hierarchical_intercepts_samples.gamma_0)
gamma_1_mean, gamma_1_var = mean_and_var(
hierarchical_intercepts_samples.gamma_1)
eps_a_means, eps_a_vars = mean_and_var(hierarchical_intercepts_samples.eps_a)
mu_a_means = gamma_0_mean + gamma_1_mean * log_uranium
mu_a_vars = gamma_0_var + np.square(log_uranium) * gamma_1_var
a_means = mu_a_means + eps_a_means[county]
a_stds = np.sqrt(mu_a_vars + eps_a_vars[county])
plt.figure()
plt.scatter(log_uranium, a_means, marker='.', c='C0')
xvals = np.linspace(-1, 0.8)
plt.plot(xvals,gamma_0_mean + gamma_1_mean * xvals, 'k--')
plt.xlim(-1, 0.8)
for ui, m, se in zip(log_uranium, a_means, a_stds):
plt.plot([ui, ui], [m - se, m + se], 'C1-', alpha=0.1)
plt.xlabel('County-level uranium')
plt.ylabel('Intercept estimate')
plot_hierarchical_intercepts()
Gli errori standard sulle intercettazioni sono più ristretti rispetto al modello di pooling parziale senza una covariata a livello di contea.
6.2 Correlazioni tra i livelli
In alcuni casi, avere predittori a più livelli può rivelare una correlazione tra variabili a livello individuale e residui di gruppo. Possiamo tenerne conto includendo la media dei predittori individuali come covariata nel modello per l'intercetta di gruppo.
\(\alpha_j = \gamma_0 + \gamma_1 u_j + \gamma_2 \bar{x} + \zeta_j\) Questi sono ampiamente indicati come effetti contestuali.
mpl.rc("font", size=12)
pgm = daft.PGM([10, 4.5], node_unit=1.2)
pgm.add_node(
daft.Node(
"gamma_prior",
r"$\mathcal{N}(0, 10^5)$",
1.5,
4,
fixed=True,
offset=(10, 5)))
pgm.add_node(daft.Node("gamma", r"$\gamma$", 1.5, 3.5))
pgm.add_node(daft.Node("mu_a", r"$\mu_a$", 1.5, 2.2))
pgm.add_node(
daft.Node(
"sigma_a_prior",
r"$\mathrm{HalfCauchy}(0, 5)$",
3,
4,
fixed=True,
offset=(0, 5)))
pgm.add_node(daft.Node("sigma_a", r"$\sigma_a$", 3, 3.5))
pgm.add_node(daft.Node("eps_a", r"$eps_a$", 3, 2.5, scale=1.25))
pgm.add_node(daft.Node("a", r"$a \sim \mathcal{Det}$", 1.5, 1.2, scale=1.5))
pgm.add_node(
daft.Node(
"sigma_prior",
r"$\mathrm{U}(0, 100)$",
4,
3.5,
fixed=True,
offset=(20, 5)))
pgm.add_node(daft.Node("sigma_y", r"$\sigma_y$", 4, 2.5))
pgm.add_node(daft.Node("b_prior", r"$\mathcal{N}(0, 10^5)$", 5, 3, fixed=True))
pgm.add_node(daft.Node("b", r"$b$", 5, 2))
pgm.add_node(
daft.Node(
"y_i", r"$y_i \sim \mathcal{N}$", 4, 1, scale=1.25, observed=True))
pgm.add_edge("gamma_prior", "gamma")
pgm.add_edge("sigma_a_prior", "sigma_a")
pgm.add_edge("sigma_a", "eps_a")
pgm.add_edge("gamma", "mu_a")
pgm.add_edge("mu_a", "a")
pgm.add_edge("eps_a", "a")
pgm.add_edge("b_prior", "b")
pgm.add_edge("sigma_prior", "sigma_y")
pgm.add_edge("sigma_y", "y_i")
pgm.add_edge("a", "y_i")
pgm.add_edge("b", "y_i")
pgm.add_plate(daft.Plate([0.9, 2.9, 1.2, 1.0], "$i = 1:3$"))
pgm.add_plate(daft.Plate([2.4, 1.7, 1.2, 1.4], "$i = 1:85$"))
pgm.add_plate(daft.Plate([0.9, 0.4, 1.2, 2.2], "$i = 1:919$"))
pgm.add_plate(daft.Plate([3.4, 0.2, 1.2, 1.4], "$i = 1:919$"))
pgm.render()
plt.show()
# Create a new variable for mean of floor across counties
xbar = tf.convert_to_tensor(radon.groupby('county')['floor'].mean(), tf.float32)
xbar = tf.gather(xbar, county, axis=-1)
def contextual_effects_model(floor, county, log_uranium, xbar):
"""Creates a joint distribution for the varying slope model."""
return tfd.JointDistributionSequential([
tfd.HalfCauchy(loc=0., scale=5), # sigma_a
lambda sigma_a: tfd.MultivariateNormalDiag( # eps_a
loc=tf.zeros([num_counties]),
scale_identity_multiplier=sigma_a),
tfd.Normal(loc=0., scale=1e5), # gamma_0
tfd.Normal(loc=0., scale=1e5), # gamma_1
tfd.Normal(loc=0., scale=1e5), # gamma_2
tfd.Normal(loc=0., scale=1e5), # b
tfd.Uniform(low=0., high=100), # sigma_y
lambda sigma_y, b, gamma_2, gamma_1, gamma_0, eps_a: tfd.
MultivariateNormalDiag( # y
loc=affine(
floor, b[..., tf.newaxis],
affine(log_uranium, gamma_1[..., tf.newaxis], gamma_0[
..., tf.newaxis]) + affine(xbar, gamma_2[..., tf.newaxis]) +
tf.gather(eps_a, county, axis=-1)),
scale_identity_multiplier=sigma_y)
])
def contextual_effects_log_prob(sigma_a, eps_a, gamma_0, gamma_1, gamma_2, b,
sigma_y):
"""Computes joint log prob pinned at `log_radon`."""
return contextual_effects_model(floor, county, log_uranium, xbar).log_prob(
[sigma_a, eps_a, gamma_0, gamma_1, gamma_2, b, sigma_y, log_radon])
@tf.function
def sample_contextual_effects(num_chains, num_results, num_burnin_steps):
"""Samples from the hierarchical intercepts model."""
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=contextual_effects_log_prob,
num_leapfrog_steps=10,
step_size=0.01)
initial_state = [
tf.ones([num_chains], name='init_sigma_a'),
tf.zeros([num_chains, num_counties], name='eps_a'),
tf.zeros([num_chains], name='init_gamma_0'),
tf.zeros([num_chains], name='init_gamma_1'),
tf.zeros([num_chains], name='init_gamma_2'),
tf.zeros([num_chains], name='init_b'),
tf.ones([num_chains], name='init_sigma_y')
]
unconstraining_bijectors = [
tfb.Exp(), # sigma_a
tfb.Identity(), # eps_a
tfb.Identity(), # gamma_0
tfb.Identity(), # gamma_1
tfb.Identity(), # gamma_2
tfb.Identity(), # b
tfb.Chain([tfb.Shift(shift=50.),
tfb.Scale(scale=50.),
tfb.Tanh()]) # sigma_y
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
ContextualEffectsModel = collections.namedtuple(
'ContextualEffectsModel',
['sigma_a', 'eps_a', 'gamma_0', 'gamma_1', 'gamma_2', 'b', 'sigma_y'])
samples, acceptance_probs = sample_contextual_effects(
num_chains=4, num_results=2000, num_burnin_steps=500)
print('Acceptance Probabilities: ', acceptance_probs.numpy())
contextual_effects_samples = ContextualEffectsModel._make(samples)
Acceptance Probabilities: [0.9505 0.9595 0.951 0.9535]
for var in ['sigma_a', 'gamma_0', 'gamma_1', 'gamma_2', 'b', 'sigma_y']:
print(
'R-hat for ', var, ': ',
tfp.mcmc.potential_scale_reduction(
getattr(contextual_effects_samples, var)).numpy())
R-hat for sigma_a : 1.0709597 R-hat for gamma_0 : 1.0067923 R-hat for gamma_1 : 1.0089629 R-hat for gamma_2 : 1.0054177 R-hat for b : 1.0018929 R-hat for sigma_y : 1.0032713
for var in ['gamma_0', 'gamma_1', 'gamma_2']:
var_samples = getattr(contextual_effects_samples, var)
mean = var_samples.numpy().mean()
std = var_samples.numpy().std()
r_hat = tfp.mcmc.potential_scale_reduction(var_samples).numpy()
n_eff = tfp.mcmc.effective_sample_size(var_samples).numpy().sum()
print(var, ' mean: ', mean, ' std: ', std, ' n_eff: ', n_eff, ' r_hat: ',
r_hat)
gamma_0 mean: 1.3934746 std: 0.04966602 n_eff: 816.21265 r_hat: 1.0067923 gamma_1 mean: 0.7229424 std: 0.088611916 n_eff: 1462.486 r_hat: 1.0089629 gamma_2 mean: 0.40893936 std: 0.20304097 n_eff: 457.8165 r_hat: 1.0054177
Quindi, potremmo dedurre da ciò che le contee con proporzioni più elevate di case senza seminterrati tendono ad avere livelli di base di radon più elevati. Forse questo è legato al tipo di terreno, che a sua volta potrebbe influenzare il tipo di strutture costruite.
6.3 Pronostico
Gelman (2006) ha utilizzato test di convalida incrociata per verificare l'errore di previsione dei modelli non raggruppati, raggruppati e parzialmente raggruppati.
Errori di previsione della convalida incrociata quadratica media radice:
- unpool = 0.86
- aggregato = 0.84
- multilivello = 0.79
Esistono due tipi di previsione che possono essere effettuati in un modello multilivello:
- Un nuovo individuo all'interno di un gruppo esistente
- Un nuovo individuo all'interno di un nuovo gruppo
Ad esempio, se volessimo fare una previsione per una nuova casa senza seminterrato nella contea di St. Louis, dobbiamo solo campionare dal modello del radon con l'intercetta appropriata.
county_name.index('St Louis')
69
Questo è,
\[\tilde{y}_i \sim N(\alpha_{69} + \beta (x_i=1), \sigma_y^2)\]
st_louis_log_uranium = tf.convert_to_tensor(
radon.where(radon['county'] == 69)['log_uranium_ppm'].mean(), tf.float32)
st_louis_xbar = tf.convert_to_tensor(
radon.where(radon['county'] == 69)['floor'].mean(), tf.float32)
@tf.function
def intercept_a(gamma_0, gamma_1, gamma_2, eps_a, log_uranium, xbar, county):
return (affine(log_uranium, gamma_1, gamma_0) + affine(xbar, gamma_2) +
tf.gather(eps_a, county, axis=-1))
def contextual_effects_predictive_model(floor, county, log_uranium, xbar,
st_louis_log_uranium, st_louis_xbar):
"""Creates a joint distribution for the contextual effects model."""
return tfd.JointDistributionSequential([
tfd.HalfCauchy(loc=0., scale=5), # sigma_a
lambda sigma_a: tfd.MultivariateNormalDiag( # eps_a
loc=tf.zeros([num_counties]),
scale_identity_multiplier=sigma_a),
tfd.Normal(loc=0., scale=1e5), # gamma_0
tfd.Normal(loc=0., scale=1e5), # gamma_1
tfd.Normal(loc=0., scale=1e5), # gamma_2
tfd.Normal(loc=0., scale=1e5), # b
tfd.Uniform(low=0., high=100), # sigma_y
# y
lambda sigma_y, b, gamma_2, gamma_1, gamma_0, eps_a: (
tfd.MultivariateNormalDiag(
loc=affine(
floor, b[..., tf.newaxis],
intercept_a(gamma_0[..., tf.newaxis],
gamma_1[..., tf.newaxis], gamma_2[..., tf.newaxis],
eps_a, log_uranium, xbar, county)),
scale_identity_multiplier=sigma_y)),
# stl_pred
lambda _, sigma_y, b, gamma_2, gamma_1, gamma_0, eps_a: tfd.Normal(
loc=intercept_a(gamma_0, gamma_1, gamma_2, eps_a,
st_louis_log_uranium, st_louis_xbar, 69) + b,
scale=sigma_y)
])
@tf.function
def contextual_effects_predictive_log_prob(sigma_a, eps_a, gamma_0, gamma_1,
gamma_2, b, sigma_y, stl_pred):
"""Computes joint log prob pinned at `log_radon`."""
return contextual_effects_predictive_model(floor, county, log_uranium, xbar,
st_louis_log_uranium,
st_louis_xbar).log_prob([
sigma_a, eps_a, gamma_0,
gamma_1, gamma_2, b, sigma_y,
log_radon, stl_pred
])
@tf.function
def sample_contextual_effects_predictive(num_chains, num_results,
num_burnin_steps):
"""Samples from the contextual effects predictive model."""
hmc = tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=contextual_effects_predictive_log_prob,
num_leapfrog_steps=50,
step_size=0.01)
initial_state = [
tf.ones([num_chains], name='init_sigma_a'),
tf.zeros([num_chains, num_counties], name='eps_a'),
tf.zeros([num_chains], name='init_gamma_0'),
tf.zeros([num_chains], name='init_gamma_1'),
tf.zeros([num_chains], name='init_gamma_2'),
tf.zeros([num_chains], name='init_b'),
tf.ones([num_chains], name='init_sigma_y'),
tf.zeros([num_chains], name='init_stl_pred')
]
unconstraining_bijectors = [
tfb.Exp(), # sigma_a
tfb.Identity(), # eps_a
tfb.Identity(), # gamma_0
tfb.Identity(), # gamma_1
tfb.Identity(), # gamma_2
tfb.Identity(), # b
tfb.Chain([tfb.Shift(shift=50.),
tfb.Scale(scale=50.),
tfb.Tanh()]), # sigma_y
tfb.Identity(), # stl_pred
]
kernel = tfp.mcmc.TransformedTransitionKernel(
inner_kernel=hmc, bijector=unconstraining_bijectors)
samples, kernel_results = tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=initial_state,
kernel=kernel)
acceptance_probs = tf.reduce_mean(
tf.cast(kernel_results.inner_results.is_accepted, tf.float32), axis=0)
return samples, acceptance_probs
ContextualEffectsPredictiveModel = collections.namedtuple(
'ContextualEffectsPredictiveModel', [
'sigma_a', 'eps_a', 'gamma_0', 'gamma_1', 'gamma_2', 'b', 'sigma_y',
'stl_pred'
])
samples, acceptance_probs = sample_contextual_effects_predictive(
num_chains=4, num_results=2000, num_burnin_steps=500)
print('Acceptance Probabilities: ', acceptance_probs.numpy())
contextual_effects_pred_samples = ContextualEffectsPredictiveModel._make(
samples)
Acceptance Probabilities: [0.9165 0.978 0.9755 0.9785]
for var in [
'sigma_a', 'gamma_0', 'gamma_1', 'gamma_2', 'b', 'sigma_y', 'stl_pred'
]:
print(
'R-hat for ', var, ': ',
tfp.mcmc.potential_scale_reduction(
getattr(contextual_effects_pred_samples, var)).numpy())
R-hat for sigma_a : 1.0325582 R-hat for gamma_0 : 1.0033548 R-hat for gamma_1 : 1.0011047 R-hat for gamma_2 : 1.001153 R-hat for b : 1.0020066 R-hat for sigma_y : 1.0128921 R-hat for stl_pred : 1.0058256
plot_traces('stl_pred', contextual_effects_pred_samples.stl_pred, num_chains=4)
plot_posterior('stl_pred', contextual_effects_pred_samples.stl_pred)
7 Conclusioni
Vantaggi dei modelli multilivello:
- Contabilità della struttura gerarchica naturale dei dati osservativi.
- Stima dei coefficienti per gruppi (sottorappresentati).
- Incorporare informazioni a livello individuale e di gruppo durante la stima dei coefficienti a livello di gruppo.
- Consentire la variazione tra i coefficienti a livello individuale tra i gruppi.
Riferimenti
Gelman, A., & Hill, J. (2006). Analisi dei dati mediante regressione e modelli multilivello/gerarchici (1a ed.). Cambridge University Press.
Gelman, A. (2006). Modellazione multilivello (gerarchica): cosa può e cosa non può fare. Tecnometria, 48(3), 432–435.