TensorFlow Probability 中的多层建模入门

本示例移植自 PyMC3 示例笔记本贝叶斯多层次建模方法入门

View on TensorFlow.org 在 Google Colab 中运行 在 Github 上查看源代码 下载笔记本

依赖项和前提条件

Import

1 简介

在本 Colab 中,我们将使用热门的 Radon 数据集拟合具有不同模型复杂度的分层线性模型 (HLM)。我们将使用 TFP 基元及其马尔可夫链蒙特卡洛工具集。

为了更好地拟合数据,我们的目标是利用数据集中存在的自然分层结构。我们从传统方式开始:完全池化和未池化的模型。我们继续处理多层次模型:探索部分池化模型、群体层次预测因子和上下文效应。

有关同样在氡数据集上使用 TFP 来拟合 HLM 的相关笔记本,请参阅 {TF Probability, R, Stan} 中的线性混合效应回归

如果对本文介绍的内容有任何疑问,请随时联系(或加入)TensorFlow Probability 邮寄名单。我们非常乐意为您提供帮助。

2 多层次建模概述

贝叶斯多层建模方法入门

分层或多层建模是回归建模的泛化。

多层次模型是回归模型,其中为组成模型参数指定了概率分布。这意味着允许模型参数按组变化。观测单元通常是自然聚类的。尽管簇本身和簇内均采用随机抽样,但聚类仍会引起观测值之间的依赖。

分层模型是一种特定的多层次模型,其中的参数彼此嵌套。一些多层次结构不是分层的。

例如,“国家”和“年份”不是嵌套的,而是可以表示单独但重叠的参数簇。我们将使用环境流行病学示例来介绍这一主题。

示例:氡污染(Gelman 和 Hill,2006 年)

氡是一种放射性气体,它通过与地面的接触点进入房屋。它是一种致癌物质,是非吸烟者患上肺癌的主要原因。不同房屋的氡水平差异很大。

EPA 对 80,000 栋房屋内的氡水平进行了研究。两个重要的预测因子分别为:1. 在地下室或一楼测量(地下室中的氡水平更高)2. 县铀水平(与氡水平呈正相关)

我们将重点研究明尼苏达州的氡水平。本例中的分层结构是每个县内的房屋。

3 数据整理

在本部分中,我们获得 radon 数据集并进行一些最少的预处理。

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()

氡水平分布(对数尺度):

plt.hist(log_radon.numpy(), bins=25, edgecolor='white')
plt.xlabel("Histogram of Radon levels (Log Scale)")
plt.show()

png

4 传统方式

对氡暴露建模的两种传统方式代表了偏差-方差权衡的两个极端:

完全池化:

对所有县一视同仁,估计单一氡水平。\(y_i = \alpha + \beta x_i + \epsilon_i\)

无池化:

在每个县对氡单独建模。

\(y_i = \alpha_{j[i]} + \beta x_i + \epsilon_i\) where \(j = 1,\ldots,85\)

误差 \(\epsilon_i\) 可以表示测量误差、房屋内部的时间变化或房屋之间的变化。

4.1 Complete Pooling Model

png

下面,我们使用汉密尔顿蒙特卡洛方法拟合完全池化模型。

@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:  [0.999 0.996 0.995 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 :     1.0019042
R-hat for  beta :  1.0135655
R-hat for  sigma :     0.99958754
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)

对于完全池化模型,为斜率和截距的点估计值绘图。

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()

png

Utility function to plot traces of sampled variables.

for var, var_samples in pooled_samples._asdict().items():
  plot_traces(var, samples=var_samples, num_chains=4)

png

png

png

接下来,我们估计未池化模型中每个县的氡水平。

4.2 Unpooled Model

png

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:  [0.895 0.897 0.893 0.901]
R-hat for beta: 1.0052257
R-hat for sigma: 1.0035229
plot_traces(var_name='beta', samples=unpooled_samples.beta, num_chains=4)
plot_traces(var_name='sigma', samples=unpooled_samples.sigma, num_chains=4)

png

png

下面是未池化县对截距的预期值以及每个链的 95% 可信区间。我们还可以报告每个县的估计值对应的 R-hat 值。

Utility function for Forest plots.

forest_plot(
    num_chains=4,
    num_vars=num_counties,
    var_name='alpha',
    var_labels=county_name,
    samples=unpooled_samples.alpha.numpy())

png

我们可以为有序估计值绘图,以确定氡水平较高的县:

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()

png

Utility function to plot estimates for a sample set of counties.

下面是对代表各种样本量的一部分县的池化和未池化估计值的可视化比较。

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)

png

这些模型都不令人满意:

  • 如果我们尝试确定氡水平较高的县,那么池化便没有用。
  • 我们不相信使用极少观测值的模型产生的极端未池化估计值。

5 多层次和分层模型

池化数据时,我们会丢失不同的数据点来自不同的县这一信息。这意味着每个 radon 水平观测值都从同一个概率分布中抽样。这种模型无法学习某个群体(例如一个县)中固有抽样单元的任何变化。它只会考虑抽样方差。

png

分析未池化的数据时,我们想当然地认为它们是从单独的模型中独立抽样。与池化的情况相反,这种方式声称抽样单元之间的差异太大而无法将它们组合:

png

在分层模型中,参数被视为来自参数总体分布的样本。 因此,我们认为它们既不是完全不同也不是完全相同。这称为部分池化

png

5.1 部分池化

房屋氡数据集最简单的部分池化模型是一种简单地估计氡水平的模型,在群体层次或个体层次上都没有任何预测因子。个体层次预测因子的一个示例是数据点是来自地下室还是来自一楼。群体层次预测因子可以是全县范围内的平均铀水平。

部分池化模型代表了池化与未池化的极端值之间的折衷,近似于未池化的县估计值和池化的估计值的加权平均值(基于样本量)。

令 \(\hat{\alpha}*j\) 为县 \(j\) 的估计对数氡水平。它只是一个截距;我们现在先忽略斜率。\(n_j\) 是来自县 \(j\) 的观测值数。 \(\sigma*{\alpha}\) 和 \(\sigma_y\) 分别是参数内的方差和抽样方差。随后,部分池化模型可以假定:

\[\hat{\alpha}_j \approx \frac{(n_j/\sigma_y^2)\bar{y}_j + (1/\sigma_{\alpha}^2)\bar{y} }{(n_j/\sigma_y^2) + (1/\sigma_{\alpha}^2)}\]

使用部分池化时,我们预期以下情形:

  • 样本量较小的县的估计值将趋向于整个州的平均值。
  • 样本量较大的县的估计值将更接近未池化的县估计值。

png

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 0.978 0.987 0.987]
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     : 1.0276643
R-hat for  sigma_a  : 1.0204039
R-hat for  sigma_y  : 1.0008202
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()

png

请注意未池化和部分池化的估计值之间的差异,尤其是在较小的样本量下。前者不但更极端,而且更不精确。

5.2 变化的截距

现在,我们考虑一个更复杂的模型,此模型允许截距根据随机效应在整个县范围内变化。

\(y_i = \alpha_{j[i]} + \beta x_{i} + \epsilon_i\),其中 \(\epsilon_i \sim N(0, \sigma_y^2)\) 以及截距随机效应:\(\alpha_{j[i]} \sim N(\mu_{\alpha}, \sigma_{\alpha}^2)\)

允许观测值根据测量位置(地下室或一楼)而变化的斜率 \(\beta\) 仍然是不同县之间共享的固定效应。

与非池化模型一样,我们为每个县设置单独的截距,而不是为每个县拟合单独的最小二乘回归模型,多层次建模在各县之间共享优势,从而可以在数据较少的县中实现更合理的推断。

png

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:  [0.989 0.98  0.988 0.983]
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 :  1.0196627
R-hat for  sigma_a :  1.0671698
R-hat for  b :  1.0017126
R-hat for  sigma_y :  0.99950683
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)

png

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)

png

png

楼层系数的估计值约为 -0.69,这可以解释为没有地下室的房屋的氡水平约为有地下室的房屋的一半 (\(\exp(-0.69) = 0.50\))(考虑县的因素后)。

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:  -0.6920927  std:  0.07004689  n_eff:  430.58865  r_hat:  1.0017126
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)')

png

5.3 变化的斜率

或者,我们可以假设一个允许县根据测量位置(地下室或一楼)如何影响氡读数而变化的模型。这种情况下,会在县之间共享截距 \(\alpha\)。

$$y_i = \alpha + \beta_{j[i]} x_{i} + \epsilon_i$$

png

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:  [0.98  0.982 0.986 0.988]
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     :  1.0972525
R-hat for  sigma_b  :  1.1294962
R-hat for  a    :  1.0047072
R-hat for  sigma_y  :  1.0015919
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)')

png

5.4 变化的截距和斜率

最通用的模型允许截距和斜率随着县变化:

$$y_i = \alpha_{j[i]} + \beta_{j[i]} x_{i} + \epsilon_i$$

png

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:  [0.989 0.958 0.984 0.985]
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     :  1.0002819
R-hat for  sigma_a  :  1.0014255
R-hat for  mu_b     :  1.0111941
R-hat for  sigma_b  :  1.0994663
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)')

png

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())

png

png

6 添加群体层次预测因子

多层次模型的主要优势是能够同时处理多个层次的预测因子。如果我们考虑上面的变化截距模型:

\(y_i = \alpha_{j[i]} + \beta x_{i} + \epsilon_i\) 我们可以指定另一个具有县层次协变量的回归模型,而不是用简单的随机效应来描述预期氡值的变化。在这里,我们使用被认为与氡水平有关的县铀读数 \(u_j\):

\(\alpha_j = \gamma_0 + \gamma_1 u_j + \zeta_j\)\(\zeta_j \sim N(0, \sigma_{\alpha}^2)\) 因此,我们现在合并了房屋层次的预测因子(楼层或地下室)以及县层次的预测因子(铀)。

请注意,此模型既包含每个县的指标变量,又包含县层次协变量。在经典回归中,这将导致共线性。在多层次模型中,将截距的部分池化向群体层次线性模型的期望值靠拢可以避免这种情况。

此外,群体层次的预测因子还用于减少群体层次的变化 \(\sigma_{\alpha}\)。这里的一个重要含义是,群体层次的估计值会引发更强大的池化。

6.1 Hierarchical Intercepts model

png

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:  [0.956  0.959  0.9675 0.958 ]
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 : 1.0204408
R-hat for gamma_0 : 1.0075455
R-hat for gamma_1 : 1.0054599
R-hat for b : 1.0011046
R-hat for sigma_y : 1.0004083
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()

png

与没有县层次协变量的部分池化模型相比,截距上的标准误差更窄。

6.2 层次之间的相关性

在某些情况下,在多个层次上拥有预测因子可以揭示个体层次变量与群体残差之间的相关性。我们可以通过将个体预测因子的平均值作为一个协变量纳入群体截距的模型来解释这一点。

\(\alpha_j = \gamma_0 + \gamma_1 u_j + \gamma_2 \bar{x} + \zeta_j\) 这些被广泛地称为上下文效应

png

# 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_diag=sigma_a[..., tf.newaxis] * tf.ones([num_counties])),
      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_diag=sigma_y[..., tf.newaxis] * tf.ones_like(xbar))
  ])


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:  [0.948 0.952 0.956 0.953]
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 :  1.1393573
R-hat for  gamma_0 :  1.0081229
R-hat for  gamma_1 :  1.0007668
R-hat for  gamma_2 :  1.012864
R-hat for  b :  1.0019505
R-hat for  sigma_y :  1.0056173
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:  1.3939122  std:  0.051875897  n_eff:  572.4374  r_hat:  1.0081229
gamma_1  mean:  0.7207277  std:  0.090660274  n_eff:  727.2628  r_hat:  1.0007668
gamma_2  mean:  0.40686083  std:  0.20155264  n_eff:  381.74048  r_hat:  1.012864

由此,我们可以推断出,没有地下室的房屋比例较高的县往往拥有较高的氡基线水平。也许这与土壤类型有关,而土壤类型反过来又可能会影响所构建的结构类型。

6.3 预测

Gelman(2006 年)使用交叉验证测试来检查未池化模型、池化模型和部分池化模型的预测误差。

交叉验证均方根预测误差:

  • 未池化 = 0.86
  • 池化 = 0.84
  • 多层次 = 0.79

在多层次模型中可以进行两种类型的预测:

  1. 现有群体中的新个体
  2. 新群体中的新个体

例如,如果我们要对圣路易斯县一所没有地下室的新房屋进行预测,我们只需要使用适当的截距从氡模型抽样。

county_name.index('St Louis')
69

即,

\[\tilde{y}_i \sim N(\alpha_{69} + \beta (x_i=1), \sigma_y^2)\]

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_diag=sigma_a[..., tf.newaxis] * tf.ones([num_counties])),
      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_diag=sigma_y[..., tf.newaxis] * tf.ones_like(xbar))),
      # 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:  [0.981  0.9795 0.972  0.9705]
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 :  1.0053602
R-hat for  gamma_0 :  1.0008001
R-hat for  gamma_1 :  1.0015156
R-hat for  gamma_2 :  0.99972683
R-hat for  b :  1.0045198
R-hat for  sigma_y :  1.0114483
R-hat for  stl_pred :  1.0045049
plot_traces('stl_pred', contextual_effects_pred_samples.stl_pred, num_chains=4)

png

plot_posterior('stl_pred', contextual_effects_pred_samples.stl_pred)

png

7 结论

多层次模型的优点:

  • 解释观测数据的自然分层结构。
  • 估算(代表性不足的)群体的系数。
  • 在估计群体层次系数时,合并个体和群体层次信息。
  • 允许各群体中的个体层次系数之间存在差异。

参考文献

Gelman, A., & Hill, J. (2006). Data Analysis Using Regression and Multilevel/Hierarchical Models (1st ed.). Cambridge University Press.

Gelman, A. (2006). Multilevel (Hierarchical) modeling: what it can and cannot do. Technometrics, 48(3), 432–435.