35  分层正态模型

This is a bit like asking how should I tweak my sailboat so I can explore the ocean floor.

— Roger Koenker 1

乔治·博克斯说,所有的模型都是错的,但有些是有用的。在真实的数据面前,尽我们所能,结果发现没有最好的模型,只有更好的模型。总是需要自己去构造符合自己需求的模型及其实现,只有自己能够实现,才能在模型的海洋中畅快地遨游。

介绍分层正态模型的定义、结构、估计,分层正态模型与曲线生长模型的关系,分层正态模型与潜变量模型的关系,分层正态模型与线性混合效应的关系。以 rstan 包和 nlme 包拟合分层正态模型,说明 rstan 包的一些用法,比较贝叶斯和频率派方法拟合的结果,给出结果的解释。再对比 16 个不同的 R包实现,总结一般地使用经验,也体会不同 R 包的独特性。

library(StanHeaders)
library(ggplot2)
library(rstan)
# 将编译的 Stan 模型与代码文件放在一起
rstan_options(auto_write = TRUE)
# 如果CPU和内存足够,设置成与马尔科夫链一样多
options(mc.cores = max(c(floor(parallel::detectCores() / 2), 1L)))
# 调色板
custom_colors <- c(
  "#4285f4", # GoogleBlue
  "#34A853", # GoogleGreen
  "#FBBC05", # GoogleYellow
  "#EA4335"  # GoogleRed
)
rstan_ggtheme_options(
  panel.background = element_rect(fill = "white"),
  legend.position = "top"
)
rstan_gg_options(
  fill = "#4285f4", color = "white",
  pt_color = "#EA4335", chain_colors = custom_colors
)
library(bayesplot)

35.1 rstan 包

本节以 8schools 数据为例介绍分层正态模型及 rstan 包实现,8schools 数据最早来自 Rubin (1981) ,分层正态模型如下:

\[ \begin{aligned} y_j &\sim \mathcal{N}(\theta_j,\sigma_j^2) \quad \theta_j = \mu + \tau \times \eta_j \\ \theta_j &\sim \mathcal{N}(\mu, \tau^2) \quad \eta_j \sim \mathcal{N}(0,1) \\ \mu &\sim \mathcal{N}(0, 100^2) \quad \tau \sim \mathrm{half\_normal}(0,100^2) \end{aligned} \]

其中,\(y_j,\sigma_j\) 是已知的观测数据,\(\theta_j\) 是模型参数, \(\eta_j\) 是服从标准正态分布的潜变量,\(\mu,\tau\) 是超参数,分别服从正态分布(将方差设置为很大的数,则变成弱信息先验或无信息均匀先验)和半正态分布(随机变量限制为正值)。

35.1.1 拟合模型

rstan 包来拟合模型,下面采用非中心的参数化表示,降低参数的相关性,减少发散的迭代次数,提高采样效率。

# 编译模型
eight_schools_fit <- stan(
  model_name = "eight_schools",
  # file = "code/eight_schools.stan",
  model_code = "
  // saved as eight_schools.stan
  data {
    int<lower=0> J;                // number of schools
    array[J] real y;               // estimated treatment effects
    array[J] real <lower=0> sigma; // standard error of effect estimates
  }
  parameters {
    real mu;                // population treatment effect
    real<lower=0> tau;      // standard deviation in treatment effects
    vector[J] eta;          // unscaled deviation from mu by school
  }
  transformed parameters {
    vector[J] theta = mu + tau * eta;        // school treatment effects
  }
  model {
    target += normal_lpdf(mu | 0, 100); 
    target += normal_lpdf(tau | 0, 100);
    target += normal_lpdf(eta | 0, 1);  // prior log-density
    target += normal_lpdf(y | theta, sigma); // log-likelihood
  }
  ",
  data = list( # 观测数据
    J = 8,
    y = c(28, 8, -3, 7, -1, 1, 18, 12),
    sigma = c(15, 10, 16, 11, 9, 11, 10, 18)
  ),
  warmup = 1000, # 每条链预处理迭代次数
  iter = 2000,   # 每条链总迭代次数
  chains = 2,    # 马尔科夫链的数目
  cores = 2,     # 指定 CPU 核心数,可以给每条链分配一个
  verbose = FALSE, # 不显示迭代的中间过程
  refresh = 0,     # 不显示采样的进度
  seed = 20232023  # 设置随机数种子,不要使用 set.seed() 函数
)
#> Warning: There were 5 divergent transitions after warmup. See
#> https://mc-stan.org/misc/warnings.html#divergent-transitions-after-warmup
#> to find out why this is a problem and how to eliminate them.
#> Warning: Examine the pairs() plot to diagnose sampling problems
#> Warning: Tail Effective Samples Size (ESS) is too low, indicating posterior variances and tail quantiles may be unreliable.
#> Running the chains for more iterations may help. See
#> https://mc-stan.org/misc/warnings.html#tail-ess

35.1.2 模型输出

用函数 print() 打印输出结果,保留 2 位小数。

print(eight_schools_fit, digits = 2)
#> Inference for Stan model: eight_schools.
#> 2 chains, each with iter=2000; warmup=1000; thin=1; 
#> post-warmup draws per chain=1000, total post-warmup draws=2000.
#> 
#>            mean se_mean   sd   2.5%    25%    50%    75%  97.5% n_eff Rhat
#> mu         7.42    0.48 5.81  -4.69   4.53   7.66  11.03  17.32   146 1.01
#> tau        6.79    0.58 6.66   0.17   2.34   5.24   8.96  25.21   132 1.02
#> eta[1]     0.37    0.02 0.93  -1.47  -0.25   0.41   1.02   2.06  1486 1.00
#> eta[2]     0.04    0.02 0.88  -1.69  -0.55   0.04   0.63   1.78  1853 1.01
#> eta[3]    -0.17    0.02 0.90  -1.92  -0.76  -0.19   0.45   1.62  1940 1.00
#> eta[4]    -0.01    0.02 0.87  -1.74  -0.59  -0.02   0.60   1.66  1356 1.00
#> eta[5]    -0.32    0.02 0.88  -2.08  -0.89  -0.34   0.25   1.50  1395 1.00
#> eta[6]    -0.17    0.02 0.86  -1.93  -0.76  -0.16   0.38   1.60  1709 1.00
#> eta[7]     0.35    0.02 0.89  -1.56  -0.24   0.40   0.93   2.04  1492 1.00
#> eta[8]     0.09    0.02 0.93  -1.73  -0.50   0.09   0.70   1.98  1759 1.00
#> theta[1]  11.02    0.21 8.09  -3.03   5.81  10.15  14.95  31.74  1421 1.00
#> theta[2]   7.86    0.14 6.46  -4.79   3.84   7.74  11.85  20.63  2055 1.00
#> theta[3]   5.86    0.45 8.41 -12.96   1.93   6.68  10.80  19.85   343 1.01
#> theta[4]   7.61    0.14 6.39  -4.98   3.58   7.78  11.50  20.37  2024 1.00
#> theta[5]   5.00    0.20 6.51  -9.71   1.29   5.59   9.34  16.59  1090 1.00
#> theta[6]   6.06    0.27 6.80  -9.09   2.34   6.52  10.53  17.94   638 1.00
#> theta[7]  10.53    0.18 6.63  -1.23   5.88  10.03  14.56  25.23  1367 1.00
#> theta[8]   8.27    0.23 7.72  -7.24   3.55   8.32  12.76  23.96  1104 1.00
#> lp__     -50.59    0.11 2.62 -56.71 -52.14 -50.38 -48.76 -46.20   586 1.00
#> 
#> Samples were drawn using NUTS(diag_e) at Sun Dec 10 15:14:03 2023.
#> For each parameter, n_eff is a crude measure of effective sample size,
#> and Rhat is the potential scale reduction factor on split chains (at 
#> convergence, Rhat=1).

值得一提,数据有限而且规律不明确,数据隐含的信息不是很多,则先验分布的情况将会对参数估计结果产生很大影响。Stan 默认采用无信息的先验分布,当使用非常弱的信息先验时,结果就非常不同了。提取任意一个参数的结果,如查看参数 \(\tau\) 的 95% 置信区间。

print(eight_schools_fit, pars = "tau", probs = c(0.025, 0.975))
#> Inference for Stan model: eight_schools.
#> 2 chains, each with iter=2000; warmup=1000; thin=1; 
#> post-warmup draws per chain=1000, total post-warmup draws=2000.
#> 
#>     mean se_mean   sd 2.5% 97.5% n_eff Rhat
#> tau 6.79    0.58 6.66 0.17 25.21   132 1.02
#> 
#> Samples were drawn using NUTS(diag_e) at Sun Dec 10 15:14:03 2023.
#> For each parameter, n_eff is a crude measure of effective sample size,
#> and Rhat is the potential scale reduction factor on split chains (at 
#> convergence, Rhat=1).

从迭代抽样数据获得与 print(fit) 一样的结果。以便后续对原始采样数据做任意的进一步分析。rstan 包扩展泛型函数 summary() 以支持对 stanfit 数据对象汇总,输出各个参数分链条和合并链条的后验分布结果。

35.1.3 操作数据

抽取数据对象 eight_schools_fit 中的采样数据,合并几条马氏链的结果,返回的结果是一个列表。

eight_schools_sim <- extract(eight_schools_fit, permuted = TRUE)

返回列表中的每个元素是一个数组,标量参数对应一维数组,向量参数对应二维数组。

str(eight_schools_sim)
#> List of 5
#>  $ mu   : num [1:2000(1d)] 5.75 7.63 19.11 16.11 10.39 ...
#>   ..- attr(*, "dimnames")=List of 1
#>   .. ..$ iterations: NULL
#>  $ tau  : num [1:2000(1d)] 10.16 8.47 17.3 4.1 6.47 ...
#>   ..- attr(*, "dimnames")=List of 1
#>   .. ..$ iterations: NULL
#>  $ eta  : num [1:2000, 1:8] 1.556 1.356 0.369 -0.433 1.293 ...
#>   ..- attr(*, "dimnames")=List of 2
#>   .. ..$ iterations: NULL
#>   .. ..$           : NULL
#>  $ theta: num [1:2000, 1:8] 21.6 19.1 25.5 14.3 18.8 ...
#>   ..- attr(*, "dimnames")=List of 2
#>   .. ..$ iterations: NULL
#>   .. ..$           : NULL
#>  $ lp__ : num [1:2000(1d)] -47.8 -47.4 -48.5 -52.1 -48.1 ...
#>   ..- attr(*, "dimnames")=List of 1
#>   .. ..$ iterations: NULL

对于列表,适合用函数 lapply() 配合算术函数计算 \(\mu,\tau\) 等参数的均值。

fun_mean <- function(x) {
  if (length(dim(x)) > 1) {
    apply(x, 2, mean)
  } else {
    mean(x)
  }
}
lapply(eight_schools_sim, FUN = fun_mean)
#> $mu
#> [1] 7.42289
#> 
#> $tau
#> [1] 6.785338
#> 
#> $eta
#> [1]  0.36786577  0.03963007 -0.16870656 -0.01330852 -0.32387749 -0.17455433
#> [7]  0.34977497  0.09338187
#> 
#> $theta
#> [1] 11.021965  7.864931  5.859438  7.610153  5.004556  6.062694 10.531080
#> [8]  8.272096
#> 
#> $lp__
#> [1] -50.59453

类似地,计算 \(\mu,\tau\) 等参数的分位点。

fun_quantile <- function(x, probs) {
  if (length(dim(x)) > 1) {
    t(apply(x, 2, quantile, probs = probs))
  } else {
    quantile(x, probs = probs)
  }
}
lapply(eight_schools_sim, fun_quantile, probs = c(2.5, 25, 50, 75, 97.5) / 100)
#> $mu
#>      2.5%       25%       50%       75%     97.5% 
#> -4.688686  4.525320  7.660881 11.026361 17.316378 
#> 
#> $tau
#>       2.5%        25%        50%        75%      97.5% 
#>  0.1725007  2.3406296  5.2397766  8.9599336 25.2140896 
#> 
#> $eta
#>       
#>             2.5%        25%         50%       75%    97.5%
#>   [1,] -1.472725 -0.2463519  0.40944442 1.0169994 2.058549
#>   [2,] -1.692349 -0.5484956  0.03527577 0.6313437 1.779431
#>   [3,] -1.915337 -0.7614554 -0.18762414 0.4470704 1.622897
#>   [4,] -1.740432 -0.5895913 -0.01766319 0.5977618 1.658543
#>   [5,] -2.077877 -0.8882916 -0.33870235 0.2505137 1.496111
#>   [6,] -1.927826 -0.7605602 -0.16412062 0.3800892 1.598579
#>   [7,] -1.560540 -0.2353192  0.39542982 0.9304768 2.036753
#>   [8,] -1.730880 -0.5037635  0.09230484 0.6956278 1.975779
#> 
#> $theta
#>       
#>              2.5%      25%       50%       75%    97.5%
#>   [1,]  -3.030548 5.811833 10.152085 14.950004 31.74400
#>   [2,]  -4.786133 3.836380  7.742984 11.846956 20.63119
#>   [3,] -12.957206 1.925727  6.679549 10.798270 19.84788
#>   [4,]  -4.978866 3.581948  7.779776 11.496501 20.37373
#>   [5,]  -9.713078 1.293159  5.589891  9.344892 16.58744
#>   [6,]  -9.086488 2.339967  6.515114 10.527118 17.94399
#>   [7,]  -1.232794 5.879803 10.032427 14.559353 25.23035
#>   [8,]  -7.244677 3.545208  8.324813 12.764714 23.95992
#> 
#> $lp__
#>      2.5%       25%       50%       75%     97.5% 
#> -56.71080 -52.13763 -50.38239 -48.76393 -46.20350

同理,可以计算最大值 max()、最小值 min() 和中位数 median() 等。

35.1.4 采样诊断

获取马尔科夫链迭代点列数据

eight_schools_sim <- extract(eight_schools_fit, permuted = FALSE)

eight_schools_sim 是一个三维数组,1000(次迭代)* 2 (条链)* 19(个参数)。如果 permuted = TRUE 则会合并马氏链的迭代结果,变成一个列表。

# 数据类型
class(eight_schools_sim)
#> [1] "array"
# 1000(次迭代)* 2 (条链)* 19(个参数)
str(eight_schools_sim)
#>  num [1:1000, 1:2, 1:19] 8.65 12.09 6.11 6.11 17.6 ...
#>  - attr(*, "dimnames")=List of 3
#>   ..$ iterations: NULL
#>   ..$ chains    : chr [1:2] "chain:1" "chain:2"
#>   ..$ parameters: chr [1:19] "mu" "tau" "eta[1]" "eta[2]" ...

提取参数 \(\mu\) 的迭代点列,绘制迭代轨迹。

eight_schools_mu_sim <- eight_schools_sim[, , "mu"]
matplot(
  eight_schools_mu_sim, xlab = "迭代次数", ylab = expression(mu),
  type = "l", lty = "solid", col = custom_colors
)
abline(h = apply(eight_schools_mu_sim, 2, mean), col = custom_colors)
legend(
  "topleft", legend = paste("chain", 1:2), box.col = "white", 
  inset = 0.01, lty = "solid", horiz = TRUE, col = custom_colors
)
图 35.1: Base R 绘制参数 \(\mu\) 的迭代轨迹

也可以使用 rstan 包提供的函数 traceplot() 或者 stan_trace() 绘制参数的迭代轨迹图。

stan_trace(eight_schools_fit, pars = "mu") +
  labs(x = "迭代次数", y = expression(mu))
图 35.2: rstan 绘制参数 \(\mu\) 的迭代轨迹

35.1.5 后验分布

可以用函数 stan_hist()stan_dens() 绘制后验分布图。下图分别展示参数 \(\mu\)\(\tau\) 的直方图,以及二者的散点图,参数 \(\mu\) 的后验概率密度分布图。

p1 <- stan_hist(eight_schools_fit, pars = c("mu","tau"), bins = 30)
p2 <- stan_scat(eight_schools_fit, pars = c("mu","tau"), size = 1) +
  labs(x = expression(mu), y = expression(tau))
p3 <- stan_dens(eight_schools_fit, pars = "mu") + labs(x = expression(mu))
library(patchwork)
p1 / (p2 + p3)
图 35.3: rstan 包绘制后验分布图

相比于 rstan 包,bayesplot 包可视化能力更强,支持对特定的参数做变换。bayesplot 包的函数 mcmc_pairs() 以矩阵图展示多个参数的分布,下图展示参数 \(\mu\)\(\log(\tau)\) 后验分布图。但是,这些函数都固定了一些标题,不能修改。

bayesplot::mcmc_pairs(
  eight_schools_fit, pars = c("mu", "tau"), transform = list(tau = "log")
)
图 35.4: bayesplot 包绘制后验分布图

35.2 其它 R 包

35.2.1 nlme

接下来,用 nlme 包拟合模型。

# 成绩
y <- c(28, 8, -3, 7, -1, 1, 18, 12)
# 标准差
sigma <- c(15, 10, 16, 11, 9, 11, 10, 18)
# 学校编号
g <- 1:8

首先,调用 nlme 包的函数 lme() 拟合模型。

library(nlme)
fit_lme <- lme(y ~ 1, random = ~ 1 | g, weights = varFixed(~ sigma^2), method = "REML")
summary(fit_lme)
#> Linear mixed-effects model fit by REML
#>   Data: NULL 
#>        AIC      BIC    logLik
#>   60.21091 60.04864 -27.10546
#> 
#> Random effects:
#>  Formula: ~1 | g
#>         (Intercept) Residual
#> StdDev:    2.917988 0.780826
#> 
#> Variance function:
#>  Structure: fixed weights
#>  Formula: ~sigma^2 
#> Fixed effects:  y ~ 1 
#>                Value Std.Error DF  t-value p-value
#> (Intercept) 7.785729  3.368082  8 2.311621  0.0496
#> 
#> Standardized Within-Group Residuals:
#>         Min          Q1         Med          Q3         Max 
#> -1.06635035 -0.73588511 -0.02896764  0.50254917  1.62502386 
#> 
#> Number of Observations: 8
#> Number of Groups: 8

随机效应的标准差 2.917988 ,随机效应部分的估计

ranef(fit_lme)
#>   (Intercept)
#> 1  1.18135690
#> 2  0.02625714
#> 3 -0.55795543
#> 4 -0.08130333
#> 5 -1.29202240
#> 6 -0.70215328
#> 7  1.25167648
#> 8  0.17414393

类比 Stan 输出结果中的 \(\theta\) 向量,每个学校的成绩估计

7.785729 + 2.917988 * ranef(fit_lme)
#>   (Intercept)
#> 1   11.232914
#> 2    7.862347
#> 3    6.157622
#> 4    7.548487
#> 5    4.015623
#> 6    5.736854
#> 7   11.438106
#> 8    8.293879

35.2.2 lme4

接着,采用 lme4 包拟合模型,发现 lme4 包获得与 nlme 包一样的结果。

control <- lme4::lmerControl(
  check.conv.singular = "ignore",
  check.nobs.vs.nRE = "ignore",
  check.nobs.vs.nlev = "ignore"
)
fit_lme4 <- lme4::lmer(y ~ 1 + (1 | g), weights = 1 / sigma^2, control = control, REML = TRUE)
summary(fit_lme4)
#> Linear mixed model fit by REML ['lmerMod']
#> Formula: y ~ 1 + (1 | g)
#> Weights: 1/sigma^2
#> Control: control
#> 
#> REML criterion at convergence: 54.2
#> 
#> Scaled residuals: 
#>      Min       1Q   Median       3Q      Max 
#> -1.06635 -0.73589 -0.02897  0.50255  1.62502 
#> 
#> Random effects:
#>  Groups   Name        Variance Std.Dev.
#>  g        (Intercept) 8.5145   2.9180  
#>  Residual             0.6097   0.7808  
#> Number of obs: 8, groups:  g, 8
#> 
#> Fixed effects:
#>             Estimate Std. Error t value
#> (Intercept)    7.786      3.368   2.312

35.2.3 blme

下面使用 blme(Chung 等 2013)blme 包基于 lme4 包,参数估计结果完全一致。

# the mode should be at the boundary of the space.

fit_blme <- blme::blmer(
  y ~ 1 + (1 | g), control = control, REML = TRUE, 
  cov.prior = NULL, weights = 1 / sigma^2
)
summary(fit_blme)
#> Prior dev  : 0
#> 
#> Linear mixed model fit by REML ['blmerMod']
#> Formula: y ~ 1 + (1 | g)
#> Weights: 1/sigma^2
#> Control: control
#> 
#> REML criterion at convergence: 54.2
#> 
#> Scaled residuals: 
#>      Min       1Q   Median       3Q      Max 
#> -1.06635 -0.73589 -0.02897  0.50255  1.62502 
#> 
#> Random effects:
#>  Groups   Name        Variance Std.Dev.
#>  g        (Intercept) 8.5145   2.9180  
#>  Residual             0.6097   0.7808  
#> Number of obs: 8, groups:  g, 8
#> 
#> Fixed effects:
#>             Estimate Std. Error t value
#> (Intercept)    7.786      3.368   2.312

35.2.4 MCMCglmm

MCMCglmm(Hadfield 2010) 采用 MCMC 算法拟合数据。

schools <- data.frame(y = y, sigma = sigma, g = g)
schools$g <- as.factor(schools$g)
# inverse-gamma prior with scale and shape equal to 0.001
prior1 <- list(
  R = list(V = diag(schools$sigma^2), fix = 1),
  G = list(G1 = list(V = 1, nu = 0.002))
)
# 为可重复
set.seed(20232023)
# 拟合模型
fit_mcmc <- MCMCglmm::MCMCglmm(
  y ~ 1, random = ~g, rcov = ~ idh(g):units, 
  data = schools, prior = prior1, verbose = FALSE
)
# 输出结果
summary(fit_mcmc)
#> 
#>  Iterations = 3001:12991
#>  Thinning interval  = 10
#>  Sample size  = 1000 
#> 
#>  DIC: -98.07615 
#> 
#>  G-structure:  ~g
#> 
#>   post.mean  l-95% CI u-95% CI eff.samp
#> g     11.23 0.0004247    68.57    361.3
#> 
#>  R-structure:  ~idh(g):units
#> 
#>          post.mean l-95% CI u-95% CI eff.samp
#> g1.units       225      225      225        0
#> g2.units       100      100      100        0
#> g3.units       256      256      256        0
#> g4.units       121      121      121        0
#> g5.units        81       81       81        0
#> g6.units       121      121      121        0
#> g7.units       100      100      100        0
#> g8.units       324      324      324        0
#> 
#>  Location effects: y ~ 1 
#> 
#>             post.mean l-95% CI u-95% CI eff.samp pMCMC  
#> (Intercept)    7.6938  -0.5149  15.3275     1023 0.062 .
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

R-structure 表示残差方差,这是已知的参数。G-structure 表示随机截距的方差,Location effects 表示固定效应的截距。截距和 nlme 包的结果很接近。

35.2.5 cmdstanr

一般地,rstan 包使用的 stan 框架版本低于 cmdstanr 包,从 rstan 包切换到 cmdstanr 包,需要注意语法、函数的变化。rstancmdstanr 使用的 Stan 版本不同导致参数估计结果不同,结果可重复的条件非常苛刻,详见 Stan 参考手册。在都是较新的版本时,Stan 代码不需要做改动,如下:

data {
  int<lower=0> J; // 学校数目 
  array[J] real y; // 测试效果的预测值
  array[J] real <lower=0> sigma; // 测试效果的标准差 
}
parameters {
  real mu; 
  real<lower=0> tau;
  vector[J] eta;
}
transformed parameters {
  vector[J] theta;
  theta = mu + tau * eta;
}
model {
  target += normal_lpdf(mu | 0, 100); 
  target += normal_lpdf(tau | 0, 100);
  target += normal_lpdf(eta | 0, 1);
  target += normal_lpdf(y | theta, sigma);
}

此处,给参数 \(\mu,\tau\) 添加了非常弱(模糊)的先验,结果将出现较大不同。

eight_schools_dat <- list(
  J = 8,
  y = c(28, 8, -3, 7, -1, 1, 18, 12),
  sigma = c(15, 10, 16, 11, 9, 11, 10, 18)
)
library(cmdstanr)
mod_eight_schools <- cmdstan_model(
  stan_file = "code/eight_schools.stan",
  compile = TRUE, cpp_options = list(stan_threads = TRUE)
)
fit_eight_schools <- mod_eight_schools$sample(
  data = eight_schools_dat, # 数据
  chains = 2,            # 总链条数
  parallel_chains = 2,   # 并行数目
  iter_warmup = 1000,    # 每条链预处理的迭代次数
  iter_sampling = 1000,  # 每条链采样的迭代次数
  threads_per_chain = 2, # 每条链设置 2 个线程
  seed = 20232023,       # 随机数种子
  show_messages = FALSE, # 不显示消息
  refresh = 0 # 不显示采样迭代的进度
)

结果保留 3 位有效数字,模型输出如下:

fit_eight_schools$summary(.num_args = list(sigfig = 3, notation = "dec"))
#> # A tibble: 19 × 10
#>    variable     mean   median    sd   mad      q5    q95  rhat ess_bulk ess_tail
#>    <chr>       <dbl>    <dbl> <dbl> <dbl>   <dbl>  <dbl> <dbl>    <dbl>    <dbl>
#>  1 lp__     -50.7    -50.4    2.54  2.43  -55.3   -46.9  1.00      515.     891.
#>  2 mu         7.92     7.85   4.71  4.55    0.133  15.7  1.00     1209.     723.
#>  3 tau        6.21     5.04   5.04  4.53    0.519  15.9  1.00      683.     843.
#>  4 eta[1]     0.324    0.334  0.982 0.976  -1.34    1.91 1.00     1855.    1148.
#>  5 eta[2]    -0.0394  -0.0396 0.852 0.860  -1.49    1.36 1.00     1882.    1405.
#>  6 eta[3]    -0.190   -0.193  0.947 0.911  -1.72    1.38 1.00     1734.    1242.
#>  7 eta[4]    -0.0307  -0.0261 0.879 0.865  -1.49    1.40 1.00     1590.    1209.
#>  8 eta[5]    -0.341   -0.369  0.896 0.841  -1.78    1.20 1.00     1687.    1370.
#>  9 eta[6]    -0.239   -0.233  0.904 0.897  -1.71    1.28 1.00     1588.    1111.
#> 10 eta[7]     0.321    0.345  0.909 0.895  -1.20    1.71 0.999    1738.    1196.
#> 11 eta[8]     0.0656   0.0700 0.952 0.965  -1.46    1.67 1.00     1847.    1296.
#> 12 theta[1]  10.7      9.48   8.26  6.58   -0.523  26.3  1.00     1866.    1225.
#> 13 theta[2]   7.61     7.54   6.14  5.70   -2.13   17.8  1.00     2351.    1734.
#> 14 theta[3]   6.15     6.50   7.38  5.99   -7.06   17.7  1.00     1874.    1252.
#> 15 theta[4]   7.55     7.65   6.19  5.79   -2.65   17.6  1.00     2050.    1436.
#> 16 theta[5]   5.17     5.65   6.29  5.86   -6.25   14.6  1.00     2072.    1656.
#> 17 theta[6]   6.01     6.50   6.87  5.98   -6.68   16.6  1.00     1802.    1251.
#> 18 theta[7]  10.4      9.89   6.52  5.95    0.518  22.1  1.00     1607.    1513.
#> 19 theta[8]   8.38     8.08   7.68  6.27   -3.65   21.0  1.00     1772.    1500.

模型采样过程的诊断结果如下:

fit_eight_schools$diagnostic_summary()
#> Warning: 2 of 2000 (0.0%) transitions ended with a divergence.
#> See https://mc-stan.org/misc/warnings for details.
#> $num_divergent
#> [1] 0 2
#> 
#> $num_max_treedepth
#> [1] 0 0
#> 
#> $ebfmi
#> [1] 1.0218599 0.8735072

分层模型的参数 \(\mu,\log(\tau)\) 的后验联合分布呈现经典的漏斗状。

bayesplot::mcmc_scatter(
  fit_eight_schools$draws(), pars = c("mu", "tau"), 
  transform = list(tau = "log"), size = 2
) + labs(x = "$\\mu$", y = "$\\log(\\tau)$")
图 35.5: 参数 \(\mu,\log(\tau)\) 的联合分布

对于调用 cmdstanr 包拟合的模型,适合用 bayesplot 包来可视化后验分布和诊断采样。

35.3 案例:rats 数据

rats 数据最早来自 Gelfand 等 (1990) ,记录 30 只小鼠每隔一周的重量,一共进行了 5 周。第一次记录是小鼠第 8 天的时候,第二次测量记录是第 15 天的时候,一直持续到第 36 天。下面在 R 环境中准备数据。

# 总共 30 只老鼠
N <- 30
# 总共进行 5 周
T <- 5
# 小鼠重量
y <- structure(c(
  151, 145, 147, 155, 135, 159, 141, 159, 177, 134,
  160, 143, 154, 171, 163, 160, 142, 156, 157, 152, 154, 139, 146,
  157, 132, 160, 169, 157, 137, 153, 199, 199, 214, 200, 188, 210,
  189, 201, 236, 182, 208, 188, 200, 221, 216, 207, 187, 203, 212,
  203, 205, 190, 191, 211, 185, 207, 216, 205, 180, 200, 246, 249,
  263, 237, 230, 252, 231, 248, 285, 220, 261, 220, 244, 270, 242,
  248, 234, 243, 259, 246, 253, 225, 229, 250, 237, 257, 261, 248,
  219, 244, 283, 293, 312, 272, 280, 298, 275, 297, 350, 260, 313,
  273, 289, 326, 281, 288, 280, 283, 307, 286, 298, 267, 272, 285,
  286, 303, 295, 289, 258, 286, 320, 354, 328, 297, 323, 331, 305,
  338, 376, 296, 352, 314, 325, 358, 312, 324, 316, 317, 336, 321,
  334, 302, 302, 323, 331, 345, 333, 316, 291, 324
), .Dim = c(30, 5))
# 第几天
x <- c(8.0, 15.0, 22.0, 29.0, 36.0)
xbar <- 22.0

重复测量的小鼠重量数据 rats 如下 表格 35.1 所示。

表格 35.1: 小鼠重量数据(部分)
第 8 天 第 15 天 第 22 天 第 29 天 第 36 天
1 151 199 246 283 320
2 145 199 249 293 354
3 147 214 263 312 328
4 155 200 237 272 297
5 135 188 230 280 323
6 159 210 252 298 331

小鼠重量数据的分布和变化情况见下图,由图可以假定 30 只小鼠的重量服从正态分布,而30 只小鼠的重量呈现一种线性增长趋势。

(a) 小鼠重量的分布
(b) 小鼠重量的变化
图 35.6: 30 只小鼠 5 次测量的数据

35.4 频率派方法

35.4.1 nlme

nlme 包适合长格式的数据,因此,先将小鼠数据整理成长格式。

rats_data <- data.frame(
  weight = as.vector(y), 
  rats = rep(1:30, times = 5), 
  days = rep(c(8, 15, 22, 29, 36), each = 30)
)

将 30 只小鼠的重量变化及回归曲线画出来,发现各只小鼠的回归线的斜率几乎一样,截距略有不同。不同小鼠的出生重量是不同,前面 Stan 采用变截距变斜率的混合效应模型拟合数据。

ggplot(data = rats_data, aes(x = days, y = weight)) +
  geom_point() +
  geom_smooth(formula = "y ~ x", method = "lm", se = FALSE) +
  theme_bw() +
  facet_wrap(facets = ~rats, labeller = "label_both", ncol = 6) +
  labs(x = "第几天", y = "重量")
图 35.7: 小鼠重量变化曲线

小鼠的重量随时间增长,不同小鼠的情况又会有所不同。作为一个参照,首先考虑变截距的随机效应模型。

\[ y_{ij} = \beta_0 + \beta_1 * x_j + \alpha_i + \epsilon_{ij}, \quad i = 1,2,\ldots,30. \quad j = 1,2,3,4,5 \]

其中,\(y_{ij}\) 表示第 \(i\) 只小鼠在第 \(j\) 次测量的重量,一共 30 只小鼠,共测量了 5 次。固定效应部分是 \(\beta_0\)\(\beta_1\) ,分别表示截距和斜率。随机效应部分是 \(\alpha_i\)\(\epsilon_{ij}\) ,分别服从正态分布\(\alpha_i \sim \mathcal{N}(0, \sigma^2_{\alpha})\)\(\epsilon_{ij} \sim \mathcal{N}(0, \sigma^2_{\epsilon})\)\(\sigma^2_{\alpha}\)\(\sigma^2_{\epsilon}\) 分别表示组间方差(group level)和组内方差(individual level)。

library(nlme)
rats_lme0 <- lme(data = rats_data, fixed = weight ~ days, random = ~ 1 | rats)
summary(rats_lme0)
#> Linear mixed-effects model fit by REML
#>   Data: rats_data 
#>        AIC     BIC    logLik
#>   1145.302 1157.29 -568.6508
#> 
#> Random effects:
#>  Formula: ~1 | rats
#>         (Intercept) Residual
#> StdDev:    14.03351 8.203811
#> 
#> Fixed effects:  weight ~ days 
#>                 Value Std.Error  DF  t-value p-value
#> (Intercept) 106.56762 3.0379720 119 35.07854       0
#> days          6.18571 0.0676639 119 91.41824       0
#>  Correlation: 
#>      (Intr)
#> days -0.49 
#> 
#> Standardized Within-Group Residuals:
#>        Min         Q1        Med         Q3        Max 
#> -2.7388198 -0.4770046  0.1261342  0.5634904  2.9981636 
#> 
#> Number of Observations: 150
#> Number of Groups: 30

当然,若考虑不同小鼠的生长速度不同(变化不是很大),可用变截距和变斜率的随机效应模型表示生长曲线模型,下面加载 nlme 包调用函数 lme() 拟合该模型。

library(nlme)
rats_lme <- lme(data = rats_data, fixed = weight ~ days, random = ~ days | rats)
summary(rats_lme)
#> Linear mixed-effects model fit by REML
#>   Data: rats_data 
#>        AIC      BIC    logLik
#>   1107.373 1125.357 -547.6867
#> 
#> Random effects:
#>  Formula: ~days | rats
#>  Structure: General positive-definite, Log-Cholesky parametrization
#>             StdDev     Corr  
#> (Intercept) 10.7426861 (Intr)
#> days         0.5105416 -0.159
#> Residual     6.0146565       
#> 
#> Fixed effects:  weight ~ days 
#>                 Value Std.Error  DF  t-value p-value
#> (Intercept) 106.56762 2.2976339 119 46.38146       0
#> days          6.18571 0.1055906 119 58.58204       0
#>  Correlation: 
#>      (Intr)
#> days -0.343
#> 
#> Standardized Within-Group Residuals:
#>        Min         Q1        Med         Q3        Max 
#> -2.6370819 -0.5394887  0.1187661  0.4927193  2.6090706 
#> 
#> Number of Observations: 150
#> Number of Groups: 30

模型输出结果中,固定效应中的截距项 (Intercept) 对应 106.56762,斜率 days 对应 6.18571。Stan 模型中截距参数 alpha0 的后验估计是 106.332,斜率参数 beta_c 的后验估计是 6.188。对比 Stan 和 nlme 包的拟合结果,可以发现贝叶斯和频率方法的结果是非常接近的。截距参数 alpha0 可以看作小鼠的初始(出生)重量,斜率参数 beta_c 可以看作小鼠的生长率 growth rate。

函数 lme() 的输出结果中,随机效应的随机截距标准差 10.7425835,对应 tau_alpha,表示每个小鼠的截距偏移量的波动。而随机斜率的标准差为 0.5105447,对应 tau_beta,相对随机截距标准差来说很小。残差标准差为 6.0146608,对应 tau_c,表示与小鼠无关的剩余量的波动,比如测量误差。总之,和 Stan 的结果有所不同,但相去不远。主要是前面的 Stan 模型没有考虑随机截距和随机斜率之间的相关性,这可以进一步调整 (Sorensen, Hohenstein, 和 Vasishth 2016)

# 参数的置信区间
intervals(rats_lme, level = 0.95)
#> Approximate 95% confidence intervals
#> 
#>  Fixed effects:
#>                  lower       est.      upper
#> (Intercept) 102.018074 106.567619 111.117164
#> days          5.976634   6.185714   6.394794
#> 
#>  Random Effects:
#>   Level: rats 
#>                            lower       est.      upper
#> sd((Intercept))        7.5158665 10.7426861 15.3548900
#> sd(days)               0.3659669  0.5105416  0.7122303
#> cor((Intercept),days) -0.5664773 -0.1590228  0.3109097
#> 
#>  Within-group standard error:
#>    lower     est.    upper 
#> 5.197149 6.014656 6.960757

Stan 输出中,截距项 alpha、斜率项 beta 参数的标准差分别是 tau_alphatau_beta ,残差标准差参数 tau_c 的估计为 6.1。简单起见,没有考虑截距项和斜率项的相关性,即不考虑小鼠出生时的重量和生长率的相关性,一般来说,应该是有关系的。函数 lme() 的输出结果中给出了截距项和斜率项的相关性为 -0.343,随机截距和随机斜率的相关性为 -0.159。

计算与 Stan 输出中的截距项 alpha_c 对应的量,结合函数 lme() 的输出,截距、斜率加和之后,如下

106.56762 + 6.18571 * 22
#> [1] 242.6532

值得注意,Stan 代码中对时间 days 做了中心化处理,即 \(x_t - \bar{x}\),目的是降低采样时参数 \(\alpha_i\)\(\beta_i\) 之间的相关性,而在拟合函数 lme() 中没有做处理,因此,结果无需转化,而且更容易解释。

fit_lm <- lm(weight ~ days, data = rats_data)
summary(fit_lm)
#> 
#> Call:
#> lm(formula = weight ~ days, data = rats_data)
#> 
#> Residuals:
#>     Min      1Q  Median      3Q     Max 
#> -38.253 -11.278   0.197   7.647  64.047 
#> 
#> Coefficients:
#>             Estimate Std. Error t value Pr(>|t|)    
#> (Intercept) 106.5676     3.2099   33.20   <2e-16 ***
#> days          6.1857     0.1331   46.49   <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> Residual standard error: 16.13 on 148 degrees of freedom
#> Multiple R-squared:  0.9359, Adjusted R-squared:  0.9355 
#> F-statistic:  2161 on 1 and 148 DF,  p-value: < 2.2e-16

采用简单线性模型即可获得与 nlme 包非常接近的估计结果,主要是小鼠重量的分布比较正态,且随时间的变化非常线性。

35.4.2 lavaan

lavaan(Rosseel 2012) 主要是用来拟合结构方程模型,而生长曲线模型可以放在该框架下。所以,也可以用 lavaan 包来拟合,并且,它提供的函数 growth() 可以直接拟合生长曲线模型。

library(lavaan)
# 设置矩阵 y 的列名
colnames(y) <- c("t1","t2","t3","t4","t5")
rats_growt_model <- " 
  # intercept and slope with fixed coefficients
  intercept =~ 1*t1 + 1*t2 + 1*t3 + 1*t4 + 1*t5
  days =~ 0*t1 + 1*t2 + 2*t3 + 3*t4 + 4*t5 

  # if we fix the variances to be equal, the models are now identical.
  t1 ~~ resvar*t1    
  t2 ~~ resvar*t2
  t3 ~~ resvar*t3
  t4 ~~ resvar*t4
  t5 ~~ resvar*t5
"

其中,算子符号 =~ 定义潜变量,~~ 定义残差协方差,intercept 表示截距, days 表示斜率。假定 5 次测量的测量误差(组内方差)是相同的。拟合模型的代码如下:

rats_growth_fit <- growth(rats_growt_model, data = y)

提供函数 summary() 获得模型输出,结果如下:

summary(rats_growth_fit, fit.measures = TRUE)
#> lavaan 0.6.16 ended normally after 87 iterations
#> 
#>   Estimator                                         ML
#>   Optimization method                           NLMINB
#>   Number of model parameters                        10
#>   Number of equality constraints                     4
#> 
#>   Number of observations                            30
#> 
#> Model Test User Model:
#>                                                       
#>   Test statistic                               106.203
#>   Degrees of freedom                                14
#>   P-value (Chi-square)                           0.000
#> 
#> Model Test Baseline Model:
#> 
#>   Test statistic                               247.075
#>   Degrees of freedom                                10
#>   P-value                                        0.000
#> 
#> User Model versus Baseline Model:
#> 
#>   Comparative Fit Index (CFI)                    0.611
#>   Tucker-Lewis Index (TLI)                       0.722
#> 
#> Loglikelihood and Information Criteria:
#> 
#>   Loglikelihood user model (H0)               -548.029
#>   Loglikelihood unrestricted model (H1)       -494.927
#>                                                       
#>   Akaike (AIC)                                1108.057
#>   Bayesian (BIC)                              1116.465
#>   Sample-size adjusted Bayesian (SABIC)       1097.783
#> 
#> Root Mean Square Error of Approximation:
#> 
#>   RMSEA                                          0.469
#>   90 Percent confidence interval - lower         0.388
#>   90 Percent confidence interval - upper         0.554
#>   P-value H_0: RMSEA <= 0.050                    0.000
#>   P-value H_0: RMSEA >= 0.080                    1.000
#> 
#> Standardized Root Mean Square Residual:
#> 
#>   SRMR                                           0.151
#> 
#> Parameter Estimates:
#> 
#>   Standard errors                             Standard
#>   Information                                 Expected
#>   Information saturated (h1) model          Structured
#> 
#> Latent Variables:
#>                    Estimate  Std.Err  z-value  P(>|z|)
#>   intercept =~                                        
#>     t1                1.000                           
#>     t2                1.000                           
#>     t3                1.000                           
#>     t4                1.000                           
#>     t5                1.000                           
#>   days =~                                             
#>     t1                0.000                           
#>     t2                1.000                           
#>     t3                2.000                           
#>     t4                3.000                           
#>     t5                4.000                           
#> 
#> Covariances:
#>                    Estimate  Std.Err  z-value  P(>|z|)
#>   intercept ~~                                        
#>     days              8.444    8.521    0.991    0.322
#> 
#> Intercepts:
#>                    Estimate  Std.Err  z-value  P(>|z|)
#>    .t1                0.000                           
#>    .t2                0.000                           
#>    .t3                0.000                           
#>    .t4                0.000                           
#>    .t5                0.000                           
#>     intercept       156.053    2.123   73.516    0.000
#>     days             43.300    0.727   59.582    0.000
#> 
#> Variances:
#>                    Estimate  Std.Err  z-value  P(>|z|)
#>    .t1      (rsvr)   36.176    5.393    6.708    0.000
#>    .t2      (rsvr)   36.176    5.393    6.708    0.000
#>    .t3      (rsvr)   36.176    5.393    6.708    0.000
#>    .t4      (rsvr)   36.176    5.393    6.708    0.000
#>    .t5      (rsvr)   36.176    5.393    6.708    0.000
#>     intrcpt         113.470   35.052    3.237    0.001
#>     days             12.226    4.126    2.963    0.003

输出结果显示 lavaan 包的函数 growth() 采用极大似然估计方法。协方差部分 Covariances: 随机效应中斜率和截距的协方差。截距部分 Intercepts: 对应于混合效应模型的固定效应部分。方差部分 Variances: 对应于混合效应模型的随机效应部分,包括残差方差、斜率和截距的方差。不难看出,这和前面 nlme 包的输出结果差别很大。原因是 lavaan 包将测量的次序从 0 开始计,0 代表小鼠出生后的第 8 天。也就是说,lavaan 采用的是次序标记,而不是实际数据。将测量发生的时间(第几天)换算成次序(第几次),并从 0 开始计,则函数 lme() 的输出和函数 growth() 就一致了。

# 重新组织数据
rats_data2 <- data.frame(
  weight = as.vector(y), 
  rats = rep(1:30, times = 5), 
  days = rep(c(0, 1, 2, 3, 4), each = 30)
)
# ML 方法估计模型参数
rats_lme2 <- lme(data = rats_data2, fixed = weight ~ days, random = ~ days | rats, method = "ML")
summary(rats_lme2)
#> Linear mixed-effects model fit by maximum likelihood
#>   Data: rats_data2 
#>        AIC      BIC    logLik
#>   1108.057 1126.121 -548.0287
#> 
#> Random effects:
#>  Formula: ~days | rats
#>  Structure: General positive-definite, Log-Cholesky parametrization
#>             StdDev    Corr  
#> (Intercept) 10.652384 (Intr)
#> days         3.496602 0.227 
#> Residual     6.014606       
#> 
#> Fixed effects:  weight ~ days 
#>                Value Std.Error  DF  t-value p-value
#> (Intercept) 156.0533 2.1370175 119 73.02389       0
#> days         43.3000 0.7316165 119 59.18401       0
#>  Correlation: 
#>      (Intr)
#> days 0.026 
#> 
#> Standardized Within-Group Residuals:
#>        Min         Q1        Med         Q3        Max 
#> -2.6317250 -0.5421585  0.1154359  0.4948025  2.6188192 
#> 
#> Number of Observations: 150
#> Number of Groups: 30

可以看到函数 growth() 给出的截距和斜率的协方差估计为 8.444,函数 lme() 给出对应截距和斜率的标准差分别是 10.652390 和 3.496588,它们的相关系数为 0.227,则函数 lme() 给出的协方差估计为 10.652390*3.496588*0.227 ,即 8.455,协方差估计比较一致。同理,比较两个输出结果中的其它成分,函数 growth() 给出的残差方差估计为 36.176,则残差标准差估计为 6.0146,结合函数 lme() 给出的 Random effects:Residual,结果完全一样。函数 growth() 给出的 Intercepts: 对应于函数 lme() 给出的固定效应部分,结果也是完全一样。

针对模型拟合对象 rats_growth_fit ,除了函数 summary() 可以汇总结果,lavaan 包还提供 AIC()BIC()logLik() 等函数,分别可以提取 AIC、BIC 和对数似然值, AIC()logLik() 结果与前面的函数 lme() 的输出是一样的,而 BIC() 不同。

35.4.3 lme4

当采用 lme4 包拟合数据的时候,发现输出结果与 nlme 包几乎相同。

rats_lme4 <- lme4::lmer(weight ~ days + (days | rats), data = rats_data)
summary(rats_lme4)
#> Linear mixed model fit by REML ['lmerMod']
#> Formula: weight ~ days + (days | rats)
#>    Data: rats_data
#> 
#> REML criterion at convergence: 1095.4
#> 
#> Scaled residuals: 
#>     Min      1Q  Median      3Q     Max 
#> -2.6371 -0.5395  0.1188  0.4927  2.6091 
#> 
#> Random effects:
#>  Groups   Name        Variance Std.Dev. Corr 
#>  rats     (Intercept) 115.4239 10.7435       
#>           days          0.2607  0.5106  -0.16
#>  Residual              36.1753  6.0146       
#> Number of obs: 150, groups:  rats, 30
#> 
#> Fixed effects:
#>             Estimate Std. Error t value
#> (Intercept) 106.5676     2.2978   46.38
#> days          6.1857     0.1056   58.58
#> 
#> Correlation of Fixed Effects:
#>      (Intr)
#> days -0.343

35.4.4 glmmTMB

glmmTMB 包基于 Template Model Builder (TMB) ,拟合广义线性混合效应模型,公式语法与 lme4 包一致。

rats_glmmtmb <- glmmTMB::glmmTMB(weight ~ days + (days | rats), REML = TRUE, data = rats_data)
summary(rats_glmmtmb)
#>  Family: gaussian  ( identity )
#> Formula:          weight ~ days + (days | rats)
#> Data: rats_data
#> 
#>      AIC      BIC   logLik deviance df.resid 
#>   1107.4   1125.4   -547.7   1095.4      146 
#> 
#> Random effects:
#> 
#> Conditional model:
#>  Groups   Name        Variance Std.Dev. Corr  
#>  rats     (Intercept) 115.4195 10.7433        
#>           days          0.2607  0.5106  -0.16 
#>  Residual              36.1756  6.0146        
#> Number of obs: 150, groups:  rats, 30
#> 
#> Dispersion estimate for gaussian family (sigma^2): 36.2 
#> 
#> Conditional model:
#>             Estimate Std. Error z value Pr(>|z|)    
#> (Intercept) 106.5676     2.2977   46.38   <2e-16 ***
#> days          6.1857     0.1056   58.58   <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

结果与 nlme 包完全一样。

35.4.5 MASS

MASS 包的结果与前面完全一致。

rats_mass <- MASS::glmmPQL(
  fixed = weight ~ days, random = ~ days | rats, 
  data = rats_data, family = gaussian(), verbose = FALSE
)
summary(rats_mass)
#> Linear mixed-effects model fit by maximum likelihood
#>   Data: rats_data 
#>   AIC BIC logLik
#>    NA  NA     NA
#> 
#> Random effects:
#>  Formula: ~days | rats
#>  Structure: General positive-definite, Log-Cholesky parametrization
#>             StdDev     Corr  
#> (Intercept) 10.4941358 (Intr)
#> days         0.4994998 -0.15 
#> Residual     6.0146494       
#> 
#> Variance function:
#>  Structure: fixed weights
#>  Formula: ~invwt 
#> Fixed effects:  weight ~ days 
#>                 Value Std.Error  DF  t-value p-value
#> (Intercept) 106.56762 2.2742323 119 46.85872       0
#> days          6.18571 0.1045144 119 59.18527       0
#>  Correlation: 
#>      (Intr)
#> days -0.343
#> 
#> Standardized Within-Group Residuals:
#>        Min         Q1        Med         Q3        Max 
#> -2.6317005 -0.5421738  0.1154530  0.4948032  2.6187746 
#> 
#> Number of Observations: 150
#> Number of Groups: 30

35.4.6 spaMM

spaMM 包的结果与前面完全一致。

rats_spamm <- spaMM::fitme(weight ~ days + (days | rats), data = rats_data)
summary(rats_spamm)
#> formula: weight ~ days + (days | rats)
#> ML: Estimation of ranCoefs and phi by ML.
#>     Estimation of fixed effects by ML.
#> Estimation of phi by 'outer' ML, maximizing logL.
#> family: gaussian( link = identity ) 
#>  ------------ Fixed effects (beta) ------------
#>             Estimate Cond. SE t-value
#> (Intercept)  106.568   2.2591   47.17
#> days           6.186   0.1038   59.58
#>  --------------- Random effects ---------------
#> Family: gaussian( link = identity ) 
#>          --- Random-coefficients Cov matrices:
#>  Group        Term   Var.   Corr.
#>   rats (Intercept)  110.1        
#>   rats        days 0.2495 -0.1507
#> # of obs: 150; # of groups: rats, 30 
#>  -------------- Residual variance  ------------
#> phi estimate was 36.1756 
#>  ------------- Likelihood values  -------------
#>                         logLik
#> logL       (p_v(h)): -548.0287
 --------------- Random effects ---------------
Family: gaussian( link = identity ) 
         --- Random-coefficients Cov matrices:
 Group        Term   Var.   Corr.
  rats (Intercept)  110.1        
  rats        days 0.2495 -0.1507
# of obs: 150; # of groups: rats, 30 

随机效应的截距方差 110.1,斜率方差 0.2495,则标准差分别是 10.49 和 0.499,相关性为 -0.1507。

 -------------- Residual variance  ------------
phi estimate was 36.1755 

残差方差为 36.1755,则标准差为 6.0146。

35.4.7 hglm

hglm(Rönnegård, Shen, 和 Alam 2010) 可以拟合分层广义线性模型,线性混合效应模型和广义线性混合效应模型,随机效应和响应变量服从的分布可以很广泛,使用语法与 lme4 包一样。

rats_hglm <- hglm::hglm2(weight ~ days + (days | rats), data = rats_data)
summary(rats_hglm)
#> Call: 
#> hglm2.formula(meanmodel = weight ~ days + (days | rats), data = rats_data)
#> 
#> ----------
#> MEAN MODEL
#> ----------
#> 
#> Summary of the fixed effects estimates:
#> 
#>             Estimate Std. Error t-value Pr(>|t|)    
#> (Intercept) 106.5676     2.1787   48.91   <2e-16 ***
#> days          6.1857     0.1029   60.13   <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> Note: P-values are based on 103 degrees of freedom
#> 
#> Summary of the random effects estimates:
#> 
#>                     Estimate Std. Error
#> (Intercept)| rats:1  -0.1705     5.3422
#> (Intercept)| rats:2  -9.8655     5.3422
#> (Intercept)| rats:3   2.7201     5.3422
#> ...
#> NOTE: to show all the random effects, use print(summary(hglm.object), print.ranef = TRUE).
#> 
#> Summary of the random effects estimates:
#> 
#>              Estimate Std. Error
#> days| rats:1  -0.1213      0.229
#> days| rats:2   0.7260      0.229
#> days| rats:3   0.3280      0.229
#> ...
#> NOTE: to show all the random effects, use print(summary(hglm.object), print.ranef = TRUE).
#> 
#> ----------------
#> DISPERSION MODEL
#> ----------------
#> 
#> NOTE: h-likelihood estimates through EQL can be biased.
#> 
#> Dispersion parameter for the mean model:
#> [1] 37.09572
#> 
#> Model estimates for the dispersion term:
#> 
#> Link = log 
#> 
#> Effects:
#>   Estimate Std. Error 
#>     3.6135     0.1391 
#> 
#> Dispersion = 1 is used in Gamma model on deviances to calculate the standard error(s).
#> 
#> Dispersion parameter for the random effects:
#> [1] 103.4501   0.2407
#> 
#> Dispersion model for the random effects:
#> 
#> Link = log
#> 
#> Effects:
#> .|Random1 
#>   Estimate Std. Error 
#>     4.6391     0.3069 
#> 
#> .|Random2 
#>   Estimate Std. Error 
#>    -1.4241     0.2920 
#> 
#> Dispersion = 1 is used in Gamma model on deviances to calculate the standard error(s).
#> 
#> EQL estimation converged in 5 iterations.

固定效应的截距和斜率都是和 nlme 包的输出结果一致。值得注意,随机效应和模型残差都是以发散参数(Dispersion parameter)来表示的,模型残差方差为 37.09572,则标准差为 6.0906,随机效应的随机截距和随机斜率的方差分别为 103.4501 和 0.2407,则标准差分别为 10.1710 和 0.4906,这与 nlme 包的结果也是一致的。

35.4.8 mgcv

先考虑一个变截距的混合效应模型

\[ y_{ij} = \beta_0 + \beta_1 * x_j + \alpha_i + \epsilon_{ij}, \quad i = 1,2,\ldots,30. \quad j = 1,2,3,4,5 \]

假设随机效应服从独立同正态分布,等价于在似然函数中添加一个岭惩罚。广义可加模型在一定形式下和上述混合效应模型存在等价关系,在广义可加模型中,可以样条表示随机效应。mgcv 包拟合代码如下。

library(mgcv)
rats_data$rats <- as.factor(rats_data$rats)
rats_gam <- gam(weight ~ days + s(rats, bs = "re"), data = rats_data)

其中,参数取值 bs = "re" 指定样条类型,re 是 Random effects 的简写。

summary(rats_gam)
#> 
#> Family: gaussian 
#> Link function: identity 
#> 
#> Formula:
#> weight ~ days + s(rats, bs = "re")
#> 
#> Parametric coefficients:
#>              Estimate Std. Error t value Pr(>|t|)    
#> (Intercept) 106.56762    3.03797   35.08   <2e-16 ***
#> days          6.18571    0.06766   91.42   <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> Approximate significance of smooth terms:
#>           edf Ref.df     F p-value    
#> s(rats) 27.14     29 14.63  <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> R-sq.(adj) =  0.983   Deviance explained = 98.6%
#> GCV = 83.533  Scale est. = 67.303    n = 150

其中,残差的方差 Scale est. = 67.303 ,则标准差为 \(\sigma_{\epsilon} = 8.2038\) 。随机效应的标准差如下

gam.vcomp(rats_gam, rescale = TRUE)
#>  s(rats) 
#> 14.03351

rescale = TRUE 表示恢复至原数据的尺度,标准差 \(\sigma_{\alpha} = 14.033\)。可以看到,固定效应和随机效应的估计结果与 nlme 包等完全一致。若考虑变截距和变斜率的混合效应模型,拟合代码如下:

rats_gam1 <- gam(
  weight ~ days + s(rats, bs = "re") + s(rats, by = days, bs = "re"),
  data = rats_data, method = "REML"
)
summary(rats_gam1)
#> 
#> Family: gaussian 
#> Link function: identity 
#> 
#> Formula:
#> weight ~ days + s(rats, bs = "re") + s(rats, by = days, bs = "re")
#> 
#> Parametric coefficients:
#>             Estimate Std. Error t value Pr(>|t|)    
#> (Intercept) 106.5676     2.2365   47.65   <2e-16 ***
#> days          6.1857     0.1028   60.18   <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> Approximate significance of smooth terms:
#>                edf Ref.df     F p-value    
#> s(rats)      21.80     29 183.9  <2e-16 ***
#> s(rats):days 23.47     29 201.8  <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> R-sq.(adj) =  0.991   Deviance explained = 99.4%
#> -REML = 547.89  Scale est. = 36.834    n = 150

输出结果中,固定效应部分的结果和 nlme 包完全一样。

gam.vcomp(rats_gam1, rescale = TRUE)
#> 
#> Standard deviations and 0.95 confidence intervals:
#> 
#>                 std.dev     lower      upper
#> s(rats)      10.3107538 7.2978205 14.5675882
#> s(rats):days  0.4916736 0.3571229  0.6769181
#> scale         6.0691017 5.2454835  7.0220401
#> 
#> Rank: 3/3

输出结果中,依次是随机效应的截距、斜率和残差的标准差(标准偏差),和 nlme 包给出的结果非常接近。

mgcv 包还提供函数 gamm(),它将混合效应和固定效应分开,在拟合 LMM 模型时,它类似 nlme 包的函数 lme()。返回一个含有 lme 和 gam 两个元素的列表,前者包含随机效应的估计,后者是固定效应的估计,固定效应中可以添加样条(或样条表示的简单随机效益,比如本节前面提及的模型)。实际上,函数 gamm() 分别调用 nlme 包和 MASS 包来拟合 LMM 模型和 GLMM 模型。

rats_gamm <- gamm(weight ~ days, random = list(rats = ~days), method = "REML", data = rats_data)
# LME
summary(rats_gamm$lme)
#> Linear mixed-effects model fit by REML
#>   Data: strip.offset(mf) 
#>        AIC      BIC    logLik
#>   1107.373 1125.357 -547.6867
#> 
#> Random effects:
#>  Formula: ~days | rats
#>  Structure: General positive-definite, Log-Cholesky parametrization
#>             StdDev     Corr  
#> (Intercept) 10.7433332 (Intr)
#> days         0.5105577 -0.159
#> Residual     6.0146119       
#> 
#> Fixed effects:  y ~ X - 1 
#>                  Value Std.Error  DF  t-value p-value
#> X(Intercept) 106.56762 2.2977301 119 46.37952       0
#> Xdays          6.18571 0.1055931 119 58.58069       0
#>  Correlation: 
#>       X(Int)
#> Xdays -0.343
#> 
#> Standardized Within-Group Residuals:
#>        Min         Q1        Med         Q3        Max 
#> -2.6371079 -0.5394997  0.1187534  0.4927191  2.6091109 
#> 
#> Number of Observations: 150
#> Number of Groups: 30
# GAM
summary(rats_gamm$gam)
#> 
#> Family: gaussian 
#> Link function: identity 
#> 
#> Formula:
#> weight ~ days
#> 
#> Parametric coefficients:
#>             Estimate Std. Error t value Pr(>|t|)    
#> (Intercept) 106.5676     2.2977   46.38   <2e-16 ***
#> days          6.1857     0.1056   58.58   <2e-16 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> 
#> R-sq.(adj) =  0.935   
#>   Scale est. = 36.176    n = 150

35.5 贝叶斯方法

35.5.1 rstan

初始化模型参数,设置采样算法的参数。

# 迭代链
chains <- 4
# 迭代次数
iter <- 1000
# 初始值
init <- rep(list(list(
  alpha = rep(250, 30), beta = rep(6, 30),
  alpha_c = 150, beta_c = 10,
  tausq_c = 1, tausq_alpha = 1,
  tausq_beta = 1
)), chains)

接下来,基于重复测量数据,建立线性生长曲线模型:

\[ \begin{aligned} \alpha_c &\sim \mathcal{N}(0,100) \quad \beta_c \sim \mathcal{N}(0,100) \\ \tau^2_{\alpha} &\sim \mathrm{inv\_gamma}(0.001, 0.001) \\ \tau^2_{\beta} &\sim \mathrm{inv\_gamma}(0.001, 0.001) \\ \tau^2_c &\sim \mathrm{inv\_gamma}(0.001, 0.001) \\ \alpha_n &\sim \mathcal{N}(\alpha_c, \tau_{\alpha}) \quad \beta_n \sim \mathcal{N}(\beta_c, \tau_{\beta}) \\ y_{nt} &\sim \mathcal{N}(\alpha_n + \beta_n * (x_t - \bar{x}), \tau_c) \\ & n = 1,2,\ldots,N \quad t = 1,2,\ldots,T \end{aligned} \]

其中, \(\alpha_c,\beta_c,\tau_c,\tau_{\alpha},\tau_{\beta}\) 为无信息先验,\(\bar{x} = 22\) 表示第 22 天,\(N = 30\)\(T = 5\) 分别表示实验中的小鼠数量和测量次数,下面采用 Stan 编码、编译、采样和拟合模型。

rats_fit <- stan(
  model_name = "rats",
  model_code = "
  data {
    int<lower=0> N;
    int<lower=0> T;
    vector[T] x;
    matrix[N,T] y;
    real xbar;
  }
  parameters {
    vector[N] alpha;
    vector[N] beta;

    real alpha_c;
    real beta_c;          // beta.c in original bugs model

    real<lower=0> tausq_c;
    real<lower=0> tausq_alpha;
    real<lower=0> tausq_beta;
  }
  transformed parameters {
    real<lower=0> tau_c;       // sigma in original bugs model
    real<lower=0> tau_alpha;
    real<lower=0> tau_beta;

    tau_c = sqrt(tausq_c);
    tau_alpha = sqrt(tausq_alpha);
    tau_beta = sqrt(tausq_beta);
  }
  model {
    alpha_c ~ normal(0, 100);
    beta_c ~ normal(0, 100);
    tausq_c ~ inv_gamma(0.001, 0.001);
    tausq_alpha ~ inv_gamma(0.001, 0.001);
    tausq_beta ~ inv_gamma(0.001, 0.001);
    alpha ~ normal(alpha_c, tau_alpha); // vectorized
    beta ~ normal(beta_c, tau_beta);  // vectorized
    for (n in 1:N)
      for (t in 1:T)
        y[n,t] ~ normal(alpha[n] + beta[n] * (x[t] - xbar), tau_c);
  }
  generated quantities {
    real alpha0;
    alpha0 = alpha_c - xbar * beta_c;
  }
  ",
  data = list(N = N, T = T, y = y, x = x, xbar = xbar),
  chains = chains, init = init, iter = iter,   
  verbose = FALSE, refresh = 0, seed = 20190425
)

模型输出结果如下:

print(rats_fit, pars = c("alpha", "beta"), include = FALSE, digits = 1)
#> Inference for Stan model: rats.
#> 4 chains, each with iter=1000; warmup=500; thin=1; 
#> post-warmup draws per chain=500, total post-warmup draws=2000.
#> 
#>               mean se_mean   sd   2.5%    25%    50%    75%  97.5% n_eff Rhat
#> alpha_c      242.5     0.1  2.6  237.3  240.8  242.5  244.2  247.6  2200    1
#> beta_c         6.2     0.0  0.1    6.0    6.1    6.2    6.3    6.4  2134    1
#> tausq_c       37.3     0.2  5.4   28.3   33.4   36.8   40.5   49.0  1090    1
#> tausq_alpha  218.5     1.5 62.1  127.3  174.5  207.2  252.1  369.2  1787    1
#> tausq_beta     0.3     0.0  0.1    0.1    0.2    0.3    0.3    0.5  1340    1
#> tau_c          6.1     0.0  0.4    5.3    5.8    6.1    6.4    7.0  1096    1
#> tau_alpha     14.6     0.0  2.0   11.3   13.2   14.4   15.9   19.2  1871    1
#> tau_beta       0.5     0.0  0.1    0.4    0.5    0.5    0.6    0.7  1304    1
#> alpha0       106.4     0.1  3.5   99.4  104.0  106.4  108.7  113.4  2300    1
#> lp__        -438.6     0.3  6.6 -452.7 -442.9 -438.1 -434.1 -426.3   703    1
#> 
#> Samples were drawn using NUTS(diag_e) at Sun Dec 10 15:16:09 2023.
#> For each parameter, n_eff is a crude measure of effective sample size,
#> and Rhat is the potential scale reduction factor on split chains (at 
#> convergence, Rhat=1).

alpha_c 表示小鼠 5 次测量的平均重量,beta_c 表示小鼠体重的增长率,\(\alpha_i,\beta_i\) 分别表示第 \(i\) 只小鼠在第 22 天(第 3 次测量或 \(x_t = \bar{x}\) )的重量和增长率(每日增加的重量)。

对于分量众多的参数向量,比较适合用岭线图展示后验分布,下面调用 bayesplot 包绘制参数向量 \(\boldsymbol{\alpha},\boldsymbol{\beta}\) 的后验分布。

# plot(rats_fit, pars = "alpha", show_density = TRUE, ci_level = 0.8, outer_level = 0.95)
bayesplot::mcmc_areas_ridges(rats_fit, pars = paste0("alpha", "[", 1:30, "]")) +
  scale_y_discrete(labels = scales::parse_format()) 
图 35.8: 参数 \(\boldsymbol{\alpha}\) 的后验分布

参数向量 \(\boldsymbol{\alpha}\) 的后验估计可以看作 \(x_t = \bar{x}\) 时小鼠的重量,上图即为各个小鼠重量的后验分布。

# plot(rats_fit, pars = "beta", ci_level = 0.8, outer_level = 0.95)
bayesplot::mcmc_areas_ridges(rats_fit, pars = paste0("beta", "[", 1:30, "]")) +
  scale_y_discrete(labels = scales::parse_format()) 
图 35.9: 参数 \(\boldsymbol{\beta}\) 的后验分布

参数向量 \(\boldsymbol{\beta}\) 的后验估计可以看作是小鼠的重量的增长率,上图即为各个小鼠重量的增长率的后验分布。

35.5.2 cmdstanr

从 rstan 包转 cmdstanr 包是非常容易的,只要语法兼容,模型代码可以原封不动。

library(cmdstanr)
mod_rats <- cmdstan_model(
  stan_file = "code/rats.stan",
  compile = TRUE, cpp_options = list(stan_threads = TRUE)
)
fit_rats <- mod_rats$sample(
  data = list(N = N, T = T, y = y, x = x, xbar = xbar), # 数据
  chains = 2,            # 总链条数
  parallel_chains = 2,   # 并行数目
  iter_warmup = 1000,    # 每条链预处理的迭代次数
  iter_sampling = 1000,  # 每条链采样的迭代次数
  threads_per_chain = 2, # 每条链设置 2 个线程
  seed = 20232023,       # 随机数种子
  show_messages = FALSE, # 不显示消息
  adapt_delta = 0.9,     # 接受率
  refresh = 0 # 不显示采样迭代的进度
)

模型输出

# 显示除了参数 alpha 和 beta 以外的结果
vars <- setdiff(fit_rats$metadata()$stan_variables, c("alpha", "beta"))
fit_rats$summary(variables = vars)
#> # A tibble: 10 × 10
#>    variable       mean   median      sd     mad       q5      q95  rhat ess_bulk
#>    <chr>         <dbl>    <dbl>   <dbl>   <dbl>    <dbl>    <dbl> <dbl>    <dbl>
#>  1 lp__       -438.    -438.     6.82    7.26   -450.    -427.    1.00      580.
#>  2 alpha_c     243.     243.     2.68    2.53    238.     247.    1.00     2315.
#>  3 beta_c        6.19     6.19   0.106   0.106     6.02     6.36  0.999    3106.
#>  4 tausq_c      37.3     36.8    5.52    5.36     29.2     46.9   1.00     1015.
#>  5 tausq_alp…  215.     206.    61.7    56.2     136.     334.    1.00     2117.
#>  6 tausq_beta    0.271    0.258  0.0941  0.0884    0.145    0.438 1.00     1526.
#>  7 tau_c         6.09     6.06   0.447   0.442     5.40     6.85  1.00     1015.
#>  8 tau_alpha    14.5     14.3    2.03    1.99     11.7     18.3   1.00     2117.
#>  9 tau_beta      0.513    0.508  0.0880  0.0879    0.380    0.662 1.00     1526.
#> 10 alpha0      106.     106.     3.57    3.56    101.     112.    1.00     2843.
#> # ℹ 1 more variable: ess_tail <dbl>

诊断信息

fit_rats$diagnostic_summary()
#> $num_divergent
#> [1] 0 0
#> 
#> $num_max_treedepth
#> [1] 0 0
#> 
#> $ebfmi
#> [1] 0.8868693 0.7464700

35.5.3 brms

brms 包是基于 rstan 包的,基于 Stan 语言做贝叶斯推断,提供与 lme4 包一致的公式语法,且扩展了模型种类。

rats_brms <- brms::brm(weight ~ days + (days | rats), data = rats_data)
summary(rats_brms)
 Family: gaussian 
  Links: mu = identity; sigma = identity 
Formula: weight ~ days + (days | rats) 
   Data: rats_data (Number of observations: 150) 
  Draws: 4 chains, each with iter = 2000; warmup = 1000; thin = 1;
         total post-warmup draws = 4000

Group-Level Effects: 
~rats (Number of levels: 30) 
                    Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
sd(Intercept)          11.27      2.23     7.36    16.08 1.00     2172     2939
sd(days)                0.54      0.09     0.37     0.74 1.00     1380     2356
cor(Intercept,days)    -0.11      0.24    -0.53     0.39 1.00      920     1541

Population-Level Effects: 
          Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
Intercept   106.47      2.47   101.61   111.23 1.00     2173     2768
days          6.18      0.11     5.96     6.41 1.00     1617     2177

Family Specific Parameters: 
      Estimate Est.Error l-95% CI u-95% CI Rhat Bulk_ESS Tail_ESS
sigma     6.15      0.47     5.30     7.14 1.00     1832     3151

Draws were sampled using sampling(NUTS). For each parameter, Bulk_ESS
and Tail_ESS are effective sample size measures, and Rhat is the potential
scale reduction factor on split chains (at convergence, Rhat = 1).

35.5.4 rstanarm

rstanarm 包与 brms 包类似,区别是前者预编译了 Stan 模型,后者根据输入数据和模型编译即时编译,此外,后者支持的模型范围更加广泛。

library(rstanarm)
rats_rstanarm <- stan_lmer(formula = weight ~ days + (days | rats), data = rats_data)
summary(rats_rstanarm)
Model Info:
 function:     stan_lmer
 family:       gaussian [identity]
 formula:      weight ~ days + (days | rats)
 algorithm:    sampling
 sample:       4000 (posterior sample size)
 priors:       see help('prior_summary')
 observations: 150
 groups:       rats (30)

Estimates:
                                      mean    sd      10%     50%     90%  
(Intercept)                         106.575   2.236 103.789 106.559 109.415
days                                  6.187   0.111   6.048   6.185   6.329
sigma                                 6.219   0.497   5.626   6.183   6.862
Sigma[rats:(Intercept),(Intercept)] 103.927  42.705  57.329  98.128 159.086
Sigma[rats:days,(Intercept)]         -0.545   1.492  -2.361  -0.402   1.162
Sigma[rats:days,days]                 0.304   0.112   0.181   0.285   0.445

MCMC diagnostics
                                    mcse  Rhat  n_eff
(Intercept)                         0.043 1.000 2753 
days                                0.003 1.005 1694 
sigma                               0.015 1.001 1172 
Sigma[rats:(Intercept),(Intercept)] 1.140 1.000 1403 
Sigma[rats:days,(Intercept)]        0.054 1.006  772 
Sigma[rats:days,days]               0.003 1.000 1456 

For each parameter, mcse is Monte Carlo standard error, 
n_eff is a crude measure of effective sample size, 
and Rhat is the potential scale reduction factor 
on split chains (at convergence Rhat=1).

固定效应的部分,截距和斜率如下:

Estimates:
                                      mean    sd      10%     50%     90%  
(Intercept)                         106.575   2.236 103.789 106.559 109.415
days                                  6.187   0.111   6.048   6.185   6.329

模型残差的标准差 sigma、随机效应 Sigma 的随机截距的方差 103.927 、随机斜率的方差 0.304 及其协方差 -0.545。

sigma                                 6.219   0.497   5.626   6.183   6.862
Sigma[rats:(Intercept),(Intercept)] 103.927  42.705  57.329  98.128 159.086
Sigma[rats:days,(Intercept)]         -0.545   1.492  -2.361  -0.402   1.162
Sigma[rats:days,days]                 0.304   0.112   0.181   0.285   0.445

rstanarmbrms 包的结果基本一致的。

35.5.5 blme

blme(Chung 等 2013) 基于 lme4(Bates 等 2015) 拟合贝叶斯线性混合效应模型。参考前面 rstan 小节中关于模型参数的先验设置,下面将残差方差的先验设置为逆伽马分布,随机效应的协方差设置为扁平分布。发现拟合结果和 nlmelme4 包的几乎一样。

rats_blme <- blme::blmer(
  weight ~ days + (days | rats), data = rats_data,
  resid.prior = invgamma, cov.prior = NULL
)
summary(rats_blme)
#> Resid prior: invgamma(shape = 0, scale = 0, posterior.scale = var)
#> Prior dev  : 7.1328
#> 
#> Linear mixed model fit by REML ['blmerMod']
#> Formula: weight ~ days + (days | rats)
#>    Data: rats_data
#> 
#> REML criterion at convergence: 1095.4
#> 
#> Scaled residuals: 
#>     Min      1Q  Median      3Q     Max 
#> -2.6697 -0.5440  0.1202  0.4968  2.6317 
#> 
#> Random effects:
#>  Groups   Name        Variance Std.Dev. Corr 
#>  rats     (Intercept) 116.3517 10.7866       
#>           days          0.2623  0.5121  -0.16
#>  Residual              35.3891  5.9489       
#> Number of obs: 150, groups:  rats, 30
#> 
#> Fixed effects:
#>             Estimate Std. Error t value
#> (Intercept) 106.5676     2.2977   46.38
#> days          6.1857     0.1056   58.58
#> 
#> Correlation of Fixed Effects:
#>      (Intr)
#> days -0.343

lme4 包的函数 lmer() 所不同的是参数 resid.priorfixef.priorcov.prior ,它们设置参数的先验分布,其它参数的含义同 lme4 包。resid.prior = invgamma 表示残差方差参数使用逆伽马分布,cov.prior = NULL 表示随机效应的协方差参数使用扁平先验 flat priors。

35.5.6 rjags

rjags (Plummer 2021) 是 JAGS 软件的 R 语言接口,可以拟合分层正态模型,再借助 coda 包 (Plummer 等 2006) 可以分析 JAGS 返回的各项数据。

JAGS 代码和 Stan 代码有不少相似之处,最大的共同点在于以直观的统计模型的符号表示编码模型,仿照 Stan 代码, JAGS 编码的模型(BUGS 代码)如下:

model {
  alpha_c ~ dnorm(0, 1.0E-4);
  beta_c ~ dnorm(0, 1.0E-4);
  
  tau_c ~ dgamma(0.001, 0.001);
  tau_alpha ~ dgamma(0.001, 0.001);
  tau_beta ~ dgamma(0.001, 0.001);

  sigma_c <- 1.0 / sqrt(tau_c);
  sigma_alpha <- 1.0 / sqrt(tau_alpha);
  sigma_beta <- 1.0 / sqrt(tau_beta);
  
  for (n in 1:N){
      alpha[n] ~ dnorm(alpha_c, tau_alpha); 
      beta[n] ~ dnorm(beta_c, tau_beta);
    for (t in 1:T) {
      y[n,t] ~ dnorm(alpha[n] + beta[n] * (x[t] - xbar), tau_c);
    }
  }
}

转化主要集中在模型块,注意二者概率分布的名称以及参数含义对应关系,JAGS 使用 precision 而不是 standard deviation or variance,比如正态分布中的方差(标准偏差)被替换为其倒数。JAGS 可以省略类型声明(初始化模型时会补上),最后,JAGS 不支持 Stan 中的向量化操作,这种新特性是独特的。

library(rjags)
# 初始值
rats_inits <- list(
  list(".RNG.name" = "base::Marsaglia-Multicarry", 
       ".RNG.seed" = 20222022, 
       "alpha_c" = 100, "beta_c" = 6, "tau_c" = 5, "tau_alpha" = 10, "tau_beta" = 0.5),
  list(".RNG.name" = "base::Marsaglia-Multicarry", 
       ".RNG.seed" = 20232023, 
       "alpha_c" = 200, "beta_c" = 10, "tau_c" = 15, "tau_alpha" = 15, "tau_beta" = 1)
)
# 模型
rats_model <- jags.model(
  file = "code/rats.bugs",
  data = list(x = x, y = y, N = 30, T = 5, xbar = 22.0),
  inits = rats_inits, 
  n.chains = 2, quiet = TRUE
)
# burn-in
update(rats_model, n.iter = 2000)
# 抽样
rats_samples <- coda.samples(rats_model,
  variable.names = c("alpha_c", "beta_c", "sigma_alpha", "sigma_beta", "sigma_c"),
  n.iter = 4000, thin = 1
)
# 参数的后验估计
summary(rats_samples)
#> 
#> Iterations = 2001:6000
#> Thinning interval = 1 
#> Number of chains = 2 
#> Sample size per chain = 4000 
#> 
#> 1. Empirical mean and standard deviation for each variable,
#>    plus standard error of the mean:
#> 
#>                 Mean      SD Naive SE Time-series SE
#> alpha_c     242.4752 2.72749 0.030494       0.031571
#> beta_c        6.1878 0.10798 0.001207       0.001481
#> sigma_alpha  14.6233 2.05688 0.022997       0.025070
#> sigma_beta    0.5176 0.09266 0.001036       0.001741
#> sigma_c       6.0731 0.46425 0.005191       0.007984
#> 
#> 2. Quantiles for each variable:
#> 
#>                 2.5%      25%      50%      75%    97.5%
#> alpha_c     237.0333 240.6832 242.5024 244.2965 247.7816
#> beta_c        5.9785   6.1150   6.1867   6.2593   6.4035
#> sigma_alpha  11.1840  13.1802  14.4152  15.8340  19.2429
#> sigma_beta    0.3571   0.4538   0.5098   0.5734   0.7187
#> sigma_c       5.2384   5.7479   6.0455   6.3803   7.0413

输出结果与 rstan 十分一致,且采样速度极快。类似地,alpha0 = alpha_c - xbar * beta_c 可得 alpha0 = 242.4752 - 22 * 6.1878 = 106.3436。

35.5.7 MCMCglmm

同前,先考虑变截距的混合效应模型,MCMCglmm(Hadfield 2010) 给出的拟合结果与 nlme 包很接近。

## 变截距模型
prior1 <- list(
  R = list(V = 1, nu = 0.002),
  G = list(G1 = list(V = 1, nu = 0.002))
)
set.seed(20232023)
rats_mcmc1 <- MCMCglmm::MCMCglmm(
  weight ~ days, random = ~ rats,
  data = rats_data, verbose = FALSE, prior = prior1
)
summary(rats_mcmc1)
#> 
#>  Iterations = 3001:12991
#>  Thinning interval  = 10
#>  Sample size  = 1000 
#> 
#>  DIC: 1088.71 
#> 
#>  G-structure:  ~rats
#> 
#>      post.mean l-95% CI u-95% CI eff.samp
#> rats       213    108.4    336.4     1000
#> 
#>  R-structure:  ~units
#> 
#>       post.mean l-95% CI u-95% CI eff.samp
#> units     68.58    50.63    86.58     1000
#> 
#>  Location effects: weight ~ days 
#> 
#>             post.mean l-95% CI u-95% CI eff.samp  pMCMC    
#> (Intercept)   106.568  100.464  112.897     1000 <0.001 ***
#> days            6.185    6.051    6.315     1000 <0.001 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

随机效应的方差(组间方差)为 211.4 ,则标准差为 14.539。残差方差(组内方差)为 68.77,则标准差为 8.293。

再考虑变截距和斜率的混合效应模型。

## 变截距、变斜率模型
prior2 <- list(
  R = list(V = 1, nu = 0.002),
  G = list(G1 = list(V = diag(2), nu = 0.002))
)
set.seed(20232023)
rats_mcmc2 <- MCMCglmm::MCMCglmm(weight ~ days,
  random = ~ us(1 + days):rats,
  data = rats_data, verbose = FALSE, prior = prior2
)
summary(rats_mcmc2)
#> 
#>  Iterations = 3001:12991
#>  Thinning interval  = 10
#>  Sample size  = 1000 
#> 
#>  DIC: 1018.746 
#> 
#>  G-structure:  ~us(1 + days):rats
#> 
#>                              post.mean l-95% CI u-95% CI eff.samp
#> (Intercept):(Intercept).rats  124.1327  41.5313  226.059    847.2
#> days:(Intercept).rats          -0.7457  -4.3090    2.571    896.6
#> (Intercept):days.rats          -0.7457  -4.3090    2.571    896.6
#> days:days.rats                  0.2783   0.1067    0.493    786.9
#> 
#>  R-structure:  ~units
#> 
#>       post.mean l-95% CI u-95% CI eff.samp
#> units     38.14    27.07    51.08     1000
#> 
#>  Location effects: weight ~ days 
#> 
#>             post.mean l-95% CI u-95% CI eff.samp  pMCMC    
#> (Intercept)    106.40   101.70   110.78    823.3 <0.001 ***
#> days             6.19     5.99     6.41    963.4 <0.001 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1

G-structure 代表随机效应部分,R-structure 代表残差效应部分,Location effects 代表固定效应部分。MCMCglmm 包的这套模型表示术语源自商业软件 ASReml

随机截距的方差为 124.1327,标准差为 11.1415,随机斜率的方差 0.2783,标准差为 0.5275,随机截距和随机斜率的协方差 -0.7457,相关系数为 -0.1268,这与 nlme 包结果很接近。

35.5.8 INLA

同前,先考虑变截距的混合效应模型。

library(INLA)
inla.setOption(short.summary = TRUE)
# 数值稳定性考虑
rats_data$weight <- rats_data$weight / 400
# 变截距
rats_inla1 <- inla(weight ~ days + f(rats, model = "iid", n = 30), 
                  family = "gaussian", data = rats_data)
# 输出结果
summary(rats_inla1)
#> Fixed effects:
#>              mean    sd 0.025quant 0.5quant 0.975quant  mode kld
#> (Intercept) 0.266 0.008      0.252    0.266      0.281 0.266   0
#> days        0.015 0.000      0.015    0.015      0.016 0.015   0
#> 
#> Model hyperparameters:
#>                                            mean     sd 0.025quant 0.5quant
#> Precision for the Gaussian observations 2414.67 311.30    1852.63  2397.34
#> Precision for rats                       888.04 244.24     494.56   859.03
#>                                         0.975quant    mode
#> Precision for the Gaussian observations    3076.07 2367.82
#> Precision for rats                         1448.25  806.15
#> 
#>  is computed

再考虑变截距和斜率的混合效应模型。

# https://inla.r-inla-download.org/r-inla.org/doc/latent/iid.pdf
# 二维高斯随机效应的先验为 Wishart prior
rats_data$rats <- as.integer(rats_data$rats)
rats_data$slopeid <- 30 + rats_data$rats
# 变截距、变斜率
rats_inla2 <- inla(
  weight ~ 1 + days + f(rats, model = "iid2d", n = 2 * 30) + f(slopeid, days, copy = "rats"),
  data = rats_data, family = "gaussian"
)
# 输出结果
summary(rats_inla2)
#> Fixed effects:
#>              mean    sd 0.025quant 0.5quant 0.975quant  mode kld
#> (Intercept) 0.266 0.034       0.20    0.266      0.333 0.266   0
#> days        0.015 0.033      -0.05    0.015      0.080 0.015   0
#> 
#> Model hyperparameters:
#>                                             mean      sd 0.025quant 0.5quant
#> Precision for the Gaussian observations 4522.629 666.547   3334.867 4480.319
#> Precision for rats (component 1)          32.097   7.956     18.938   31.260
#> Precision for rats (component 2)          33.234   8.227     19.603   32.377
#> Rho1:2 for rats                           -0.001   0.172     -0.335   -0.001
#>                                         0.975quant     mode
#> Precision for the Gaussian observations   5952.954 4407.855
#> Precision for rats (component 1)            50.048   29.770
#> Precision for rats (component 2)            51.773   30.859
#> Rho1:2 for rats                              0.334   -0.001
#> 
#>  is computed
警告

对于变截距和斜率混合效应模型,还未完全弄清楚 INLA 包的输出结果。固定效应部分和残差部分都是和前面一致的,但不清楚随机效应的方差协方差矩阵的估计与 INLA 输出的对应关系。参考《Bayesian inference with INLA》第 3 章第 3 小节。

35.6 总结

基于 rats 数据建立变截距、变斜率的分层正态模型,也是线性混合效应模型的一种特殊情况,下表给出不同方法对模型各个参数的估计及置信区间。

表格 35.2: 频率派方法比较
\(\beta_0\) \(\beta_1\) \(\sigma_0\) \(\sigma_1\) \(\rho_{\sigma}\) \(\sigma_{\epsilon}\)
nlme (REML) 106.568 6.186 10.743 0.511 -0.159 6.015
lme4 (REML) 106.568 6.186 10.744 0.511 -0.16 6.015
glmmTMB (REML) 106.568 6.186 10.743 0.511 -0.16 6.015
MASS (PQL) 106.568 6.186 10.495 0.500 -0.15 6.015
spaMM (ML) 106.568 6.186 10.49 0.499 -0.15 6.015
hglm 106.568 6.186 10.171 0.491 - 6.091
mgcv (REML) 106.568 6.186 10.311 0.492 - 6.069

表中给出截距 \(\beta_0\) 、斜率 \(\beta_1\) 、随机截距 \(\sigma_0\)、随机斜率 \(\sigma_1\)、随机截距和斜率的相关系数 \(\rho_{\sigma}\)、残差 \(\sigma_{\epsilon}\) 等参数的估计及 95% 的置信区间,四舍五入保留 3 位小数。固定效应部分的结果完全相同,随机效应部分略有不同。

表格 35.3: 贝叶斯方法比较
\(\beta_0\) \(\beta_1\) \(\sigma_0\) \(\sigma_1\) \(\rho_{\sigma}\) \(\sigma_{\epsilon}\)
rstan (NUTS) 106.4 6.2 14.6 0.5 - 6.1
cmdstanr (NUTS) 106 6.19 14.5 0.513 - 6.09
brms (NUTS) 106.47 6.18 11.27 0.54 -0.11 6.15
rstanarm (NUTS) 106.575 6.187 10.194 0.551 -0.0969 6.219
blme (REML) 106.568 6.186 10.787 0.512 -0.160 5.949
rjags (Gibbs) 106.344 6.188 14.623 0.518 - 6.073
MCMCglmm (MCMC) 106.40 6.19 11.14 0.53 -0.13 6.18

其中,INLA 结果的转化未完成,表格中暂缺。rstancmdstanrrjags 未考虑随机截距和随机斜率的相关性,因此,相关系数暂缺。MCMC 是一种随机优化算法,在不同的实现中,可重复性的要求不同,设置随机数种子仅是其中的一个必要条件,故而,每次运行程序结果可能略微不同,但不影响结论。Stan 相关的 R 包输出结果中,rstan 保留 1 位小数,cmdstanr 保留 3 位有效数字,brms 保留 2 位小数,rstanarm 小数点后保留 3 位有效数字,各不相同,暂未统一处理。

35.7 习题

  1. 四个组的重复测量数据,如下表所示,建立贝叶斯线性混合效应模型/分层正态模型分析数据,与 nlme 包拟合的结果对比。

    表格 35.4: 实验数据
    编号 第1组 第2组 第3组 第4组
    1 62 63 68 56
    2 60 67 66 62
    3 63 71 71 60
    4 59 64 67 61
    5 65 68 63
    6 66 68 64
    7 63
    8 59

    \[ \begin{aligned} y_{ij} \sim \mathcal{N}(\theta_i, \sigma^2) &\quad \theta_i \sim \mathcal{N}(\mu, \tau^2) \\ (\mu,\log \sigma, \tau) &\sim \mathrm{uniform\ prior} \\ i = 1,2,3,4 &\quad j = 1,2, \ldots, n_i \end{aligned} \]

    \(y_{ij}\) 表示第 \(i\) 组的第 \(j\) 个测量值,\(\theta_i\) 表示第 \(i\) 组的均值,\(\mu\) 表示整体的均值,\(\sigma^2\) 表示组内的方差,\(\tau^2\) 表示组内的方差。

    library(nlme)
    fit_lme <- lme(data = dat, fixed = y ~ 1, random = ~ 1 | group)
    summary(fit_lme)
    #> Linear mixed-effects model fit by REML
    #>   Data: dat 
    #>        AIC      BIC    logLik
    #>   121.7804 125.1869 -57.89019
    #> 
    #> Random effects:
    #>  Formula: ~1 | group
    #>         (Intercept) Residual
    #> StdDev:    3.419288 2.366309
    #> 
    #> Fixed effects:  y ~ 1 
    #>                Value Std.Error DF  t-value p-value
    #> (Intercept) 64.01266  1.780313 20 35.95584       0
    #> 
    #> Standardized Within-Group Residuals:
    #>         Min          Q1         Med          Q3         Max 
    #> -2.18490896 -0.59921167  0.09332131  0.54077636  2.17507789 
    #> 
    #> Number of Observations: 24
    #> Number of Groups: 4

    随机效应(组间标准差)\(\tau^2\) 3.419288 、残差效应(组内标准差)\(\sigma^2\) 2.366309。截距 \(\mu\) 64.01266 代表整体的均值。各组的均值如下:

    64.01266 + ranef(fit_lme)
    #>   (Intercept)
    #> 1    61.32214
    #> 2    65.85309
    #> 3    67.70525
    #> 4    61.17016

    也可以调用 rjags 包连接 JAGS 软件做贝叶斯推理,JAGS 代码如下:

    model {
      ## specify the distribution for observations
      for(i in 1:n){
        y[i] ~ dnorm(theta[group[i]], 1/sigma2)
      }
    
      ## specify the prior for theta
      for(j in 1:J){
        theta[j] ~ dnorm(mu, 1/tau2)
      }
    
      ## specify the prior for hyperparameters
      mu ~ dunif(55, 75)
    
      log_sigma ~ dunif(-10, 3)
      sigma2 <- exp(2*log_sigma)
      sigma <- exp(log_sigma)
    
      tau ~ dunif(0, 8)
      tau2 <- pow(tau, 2)
    }

    完整的运行代码如下:

    library(rjags)
    # 参考值
    mu_a <- min(y)
    mu_b <- max(y)
    log_sigma_b <- 2 * log(sd(y))
    tau_b <- 2 * sd(y)
    
    J <- 4            # 4 个组
    n <- length(y)    # 观察值数量
    N <- 1500         # 总采样数
    nthin <- 1        # 采样间隔
    nchains <- 2      # 2 条链
    ndiscard <- N / 2 # 预处理阶段 warm-up / burn-in
    
    # 初始值
    jags_inits <- list(
      list(".RNG.name" = "base::Marsaglia-Multicarry", 
           ".RNG.seed" = 20222022, 
           "theta" = rep(3, 4), "mu" = 60, "log_sigma" = 0, "tau" = 1.5),
      list(".RNG.name" = "base::Marsaglia-Multicarry", 
           ".RNG.seed" = 20232023, 
           "theta" = rep(2, 4), "mu" = 60, "log_sigma" = 1, "tau" = 0.375)
    )
    # Call JAGS from R
    jags_model <- jags.model(
      file = "code/hnm.bugs",
      data = list("y" = y, "group" = group, "J" = J, "n" = n),
      inits = jags_inits, n.chains = nchains, quiet = TRUE
    )
    # burn-in
    update(jags_model, n.iter = ndiscard)
    # 抽样
    jags_samples <- coda.samples(jags_model,
      variable.names = c('theta','mu','sigma','tau'), n.iter = N
    )
    # 参数的后验估计
    summary(jags_samples)
    #> 
    #> Iterations = 1751:3250
    #> Thinning interval = 1 
    #> Number of chains = 2 
    #> Sample size per chain = 1500 
    #> 
    #> 1. Empirical mean and standard deviation for each variable,
    #>    plus standard error of the mean:
    #> 
    #>            Mean     SD Naive SE Time-series SE
    #> mu       64.142 2.3470  0.04285        0.05879
    #> sigma     2.473 0.4278  0.00781        0.01117
    #> tau       4.372 1.5995  0.02920        0.05101
    #> theta[1] 61.356 1.2301  0.02246        0.02503
    #> theta[2] 65.877 1.0056  0.01836        0.01928
    #> theta[3] 67.696 1.0247  0.01871        0.02119
    #> theta[4] 61.186 0.8694  0.01587        0.01692
    #> 
    #> 2. Quantiles for each variable:
    #> 
    #>            2.5%    25%    50%    75%  97.5%
    #> mu       59.145 62.700 64.167 65.510 69.027
    #> sigma     1.795  2.165  2.424  2.718  3.471
    #> tau       1.846  3.128  4.171  5.464  7.652
    #> theta[1] 58.947 60.545 61.342 62.161 63.771
    #> theta[2] 63.866 65.228 65.878 66.548 67.872
    #> theta[3] 65.665 67.046 67.712 68.337 69.692
    #> theta[4] 59.446 60.632 61.189 61.707 62.975
  2. 基于 lme4 包中学生对老师的评价数据 InstEval 建立(广义)线性混合效应模型分析数据。将响应变量(学生评价)视为有序的离散型变量,比较观察两个模型拟合效果(lme4、GLMMadaptive、spaMM 都不支持有序的响应变量,brms 则支持各类有序回归,使用语法与 lme4 完全一样。但是,由于数据规模比较大,计算时间数以天计,可考虑用 Stan 直接编码)。再者,从 Stan 实现的贝叶斯模型来看,感受 Stan 建模的灵活性和扩展性。(nlme 包不支持此等交叉随机效应的表达。)

    data(InstEval, package = "lme4")
    str(InstEval)
    #> 'data.frame':    73421 obs. of  7 variables:
    #>  $ s      : Factor w/ 2972 levels "1","2","3","4",..: 1 1 1 1 2 2 3 3 3 3 ...
    #>  $ d      : Factor w/ 1128 levels "1","6","7","8",..: 525 560 832 1068 62 406 3 6 19 75 ...
    #>  $ studage: Ord.factor w/ 4 levels "2"<"4"<"6"<"8": 1 1 1 1 1 1 1 1 1 1 ...
    #>  $ lectage: Ord.factor w/ 6 levels "1"<"2"<"3"<"4"<..: 2 1 2 2 1 1 1 1 1 1 ...
    #>  $ service: Factor w/ 2 levels "0","1": 1 2 1 2 1 1 2 1 1 1 ...
    #>  $ dept   : Factor w/ 14 levels "15","5","10",..: 14 5 14 12 2 2 13 3 3 3 ...
    #>  $ y      : int  5 2 5 3 2 4 4 5 5 4 ...
    • 因子型变量 s 表示 1-2972 位参与评分的学生。
    • 因子型变量 d 表示 1-2160 位上课的讲师。
    • 因子型变量 dept 表示课程相关的 1-15 院系。
    • 因子型变量 service 表示讲师除了授课外,是否承担其它服务。
    • 数值型变量 y 表示学生给课程的评分,1-5 分对应从坏到很好。
    # 数值型的响应变量
    fit_lme4 <- lme4::lmer(y ~ 1 + service + (1 | s) + (1 | d) + (1 | dept), data = InstEval)
    summary(fit_lme4)
    #> Linear mixed model fit by REML ['lmerMod']
    #> Formula: y ~ 1 + service + (1 | s) + (1 | d) + (1 | dept)
    #>    Data: InstEval
    #> 
    #> REML criterion at convergence: 237733.8
    #> 
    #> Scaled residuals: 
    #>     Min      1Q  Median      3Q     Max 
    #> -3.0597 -0.7478  0.0404  0.7723  3.1988 
    #> 
    #> Random effects:
    #>  Groups   Name        Variance Std.Dev.
    #>  s        (Intercept) 0.105998 0.32557 
    #>  d        (Intercept) 0.265221 0.51500 
    #>  dept     (Intercept) 0.006912 0.08314 
    #>  Residual             1.386500 1.17750 
    #> Number of obs: 73421, groups:  s, 2972; d, 1128; dept, 14
    #> 
    #> Fixed effects:
    #>             Estimate Std. Error t value
    #> (Intercept)  3.28259    0.02935 111.859
    #> service1    -0.09264    0.01339  -6.919
    #> 
    #> Correlation of Fixed Effects:
    #>          (Intr)
    #> service1 -0.152

    lme4 包不支持响应变量为有序分类变量的情形,可用 ordinal 包,此等规模数据,拟合模型需要 5-10 分钟时间。

    # 有序因子型的响应变量
    InstEval$y <- factor(InstEval$y, ordered = TRUE)
    library(ordinal)
    fit_ordinal <- clmm(
      y ~ 1 + service + (1 | s) + (1 | d) + (1 | dept),
      data = InstEval, link = "probit", threshold = "equidistant"
    )
    summary(fit_ordinal)
    
    ## MCMCglmm
    library(MCMCglmm)
    prior2 <- list(
      R = list(V = 1, nu = 0.002),
      G = list(
        G1 = list(V = 1, nu = 0.002),
        G2 = list(V = 1, nu = 0.002),
        G3 = list(V = 1, nu = 0.002)
      )
    )
    # 响应变量视为数值变量
    fit_mcmc2 <- MCMCglmm(
      y ~ service, random = ~ s + d + dept, family = "gaussian",
      data = InstEval, verbose = FALSE, prior = prior2
    )
    # 响应变量视为有序的分类变量
    fit_mcmc3 <- MCMCglmm(
      y ~ service, random = ~ s + d + dept, family = "ordinal",
      data = InstEval, verbose = FALSE, prior = prior2
    )

    当数据量较大时,MCMCglmm 包拟合模型需要很长时间,放弃,此时,Stan 的相对优势可以体现出来了。Stan 适合大型复杂概率统计模型。