Developing Hierarchical Version of Nonlinear Growth Curve Model in Stan

The following model is model 1 of Preece and Baines (1978, Annals of Human Biology), and is used to describe human growth.

My Stan code for this model is as follows:

```{stan output.var="test"}
 data {
  int<lower=1> n;
  ordered[n] t; // age
  ordered[n] y; // height of human
 }

 parameters {
  positive_ordered[2] h;
  real<lower=0, upper=t[n-1]>theta;
  positive_ordered[2] s; 
  real<lower=0>sigma;
 }

 model {
  h[1] ~ uniform(0, y[n]);
  h[2] ~ normal(180, 20);
  sigma ~ student_t(2, 0, 1);
  s[1] ~ normal(5, 5);
  s[2] ~ normal(5, 5);
  theta ~ normal(10, 5);

  y ~ normal(h[2] - (2*(h[2] - h[1]) * inv(exp(s[1]*(t - theta)) + exp(s[2]*(t - theta)))), sigma);
 }
```

I now want to create a hierarchical version of this model, with parameters modelled either identical across boys (such as the measurement variability sigma) or in a hierarchical version (such as h_1, the adult height).

As for the parameters sigma and theta, I want to maintain these priors as identical across boys as sigma ~ student_t(2, 0, 1); and theta ~ normal(10, 5);.

Unfortunately, I have almost no experience in implementing hierarchical modelling, and I have struggled in my attempts at doing any hierarchical examples beyond the simple binomial hierarchical models in Bayesian textbooks (see chapter 12, pages 358-359 of Statistical Rethinking by Richard McElreath). I do, however, understand the theory behind hierarchical modelling, as written in chapter 5 of Bayesian Data Analysis by Andrew Gelman and chapter 12 of Statistical Rethinking by Richard McElreath.

I would appreciate it if people could please take the time to demonstrate how this type of hierarchical model would be implemented in Stan. If it isn't too much to ask, I would greatly appreciate any explanations you could provide alongside the code, so that I may learn how to implement these types of hierarchical model examples independently in the future.

A sample of my data is as follows:

    age boy01 boy02 boy03 boy04 boy05 boy06 boy07 boy08 boy09 boy10 boy11 boy12 boy13 boy14 boy15 boy16 boy17 boy18
 1  1     81.3  76.2  76.8  74.1  74.2  76.8  72.4  73.8  75.4  78.8  76.9  81.6  78    76.4  76.4  76.2  75    79.7
 2  1.25  84.2  80.4  79.8  78.4  76.3  79.1  76    78.7  81    83.3  79.9  83.7  81.8  79.4  81.2  79.2  78.4  81.3
 3  1.5   86.4  83.2  82.6  82.6  78.3  81.1  79.4  83    84.9  87    84.1  86.3  85    83.4  86    82.3  82    83.3
 4  1.75  88.9  85.4  84.7  85.4  80.3  84.4  82    85.8  87.9  89.6  88.5  88.8  86.4  87.6  89.2  85.4  84    86.5
 5  2     91.4  87.6  86.7  88.1  82.2  87.4  84.2  88.4  90    91.4  90.6  92.2  87.1  91.4  92.2  88.4  85.9  88.9
 6  3    101.   97    94.2  98.6  89.4  94    93.2  97.3  97.3 100.   96.6  99.3  96.2 101.  101.  101    95.6  99.4
 7  4    110.  105.  100.  104.   96.9 102.  102.  107.  103.  111   105.  106.  104   106.  110.  107.  102.  104. 
 8  5    116.  112.  107.  111   104.  109.  109   113.  108.  118.  112   113.  111   113.  117.  115.  109.  112. 
 9  6    122.  119.  112.  116.  111.  116.  117.  119.  114.  126.  119.  120.  117.  120.  122.  121.  118.  119  
10  7    130   125   119.  123.  116.  122.  123.  126.  120.  131.  125.  127.  124.  129.  130.  128   125.  128

My apologies for the lack of precision in the decimal places. The data is in the form of a tibble table, which doesn't seem to respond to R's usual commands for greater precision. For the sake of consistency, it might be better to simply ignore all of the rows after row 5, since rows 1 - 5 display the full precision that is present in the original data.

In the full data, the ages are

> Children$age
 [1]  1.00  1.25  1.50  1.75  2.00  3.00  4.00  5.00  6.00  7.00  8.00  8.50  9.00  9.50 10.00 10.50 11.00 11.50 12.00 12.50
[21] 13.00 13.50 14.00 14.50 15.00 15.50 16.00 16.50 17.00 17.50 18.00

And there are 39 boys, listed in the same wide-data format as the above sample.

Disclaimer: As a start let's fit a (non-hierarchical) non-linear growth model using Stan.

We read in the sample data.

library(tidyverse);
df <- read.table(text = "
age boy01 boy02 boy03 boy04 boy05 boy06 boy07 boy08 boy09 boy10 boy11 boy12 boy13 boy14 boy15 boy16 boy17 boy18
1  1     81.3  76.2  76.8  74.1  74.2  76.8  72.4  73.8  75.4  78.8  76.9  81.6  78    76.4  76.4  76.2  75    79.7
2  1.25  84.2  80.4  79.8  78.4  76.3  79.1  76    78.7  81    83.3  79.9  83.7  81.8  79.4  81.2  79.2  78.4  81.3
3  1.5   86.4  83.2  82.6  82.6  78.3  81.1  79.4  83    84.9  87    84.1  86.3  85    83.4  86    82.3  82    83.3
4  1.75  88.9  85.4  84.7  85.4  80.3  84.4  82    85.8  87.9  89.6  88.5  88.8  86.4  87.6  89.2  85.4  84    86.5
5  2     91.4  87.6  86.7  88.1  82.2  87.4  84.2  88.4  90    91.4  90.6  92.2  87.1  91.4  92.2  88.4  85.9  88.9
6  3    101.   97    94.2  98.6  89.4  94    93.2  97.3  97.3 100.   96.6  99.3  96.2 101.  101.  101    95.6  99.4
7  4    110.  105.  100.  104.   96.9 102.  102.  107.  103.  111   105.  106.  104   106.  110.  107.  102.  104.
8  5    116.  112.  107.  111   104.  109.  109   113.  108.  118.  112   113.  111   113.  117.  115.  109.  112.
9  6    122.  119.  112.  116.  111.  116.  117.  119.  114.  126.  119.  120.  117.  120.  122.  121.  118.  119
10  7    130   125   119.  123.  116.  122.  123.  126.  120.  131.  125.  127.  124.  129.  130.  128   125.  128", header = T, row.names = 1);
df <- df %>%
    gather(boy, height, -age);

We define the Stan model.

model <- "
data {
    int N;                                       // Number of observations
    real y[N];                                   // Height
    real t[N];                                   // Time
}

parameters {
    real<lower=0,upper=1> s[2];
    real h_theta;
    real theta;
    real sigma;
}

transformed parameters {
    vector[N] mu;
    real h1;
    h1 = max(y);
    for (i in 1:N)
        mu[i] = h1 - 2 * (h1 - h_theta) / (exp(s[1] * (t[i] - theta)) + (exp(s[2] * (t[i] - theta))));
}

model {
    // Priors
    s ~ cauchy(0, 2.5);

    y ~ normal(mu, sigma);
}
"

Here we consider weakly informative (regularising) priors on s[1] and s[2].

Fit the model to the data.

library(rstan);
options(mc.cores = parallel::detectCores());
rstan_options(auto_write = TRUE);    
model <- stan(
    model_code = model,
    data = list(
        N = nrow(df),
        y = df$height,
        t = df$age),
    iter = 4000);

Summary of the 6 model parameters:

summary(model, pars = c("h1", "h_theta", "theta", "s", "sigma"))$summary 
#               mean     se_mean        sd        2.5%         25%         50%
#h1      131.0000000 0.000000000 0.0000000 131.0000000 131.0000000 131.0000000
#h_theta 121.6874553 0.118527828 2.7554944 115.4316738 121.1654809 122.2134014
#theta     6.5895553 0.019738319 0.5143429   5.4232740   6.4053479   6.6469534
#s[1]      0.7170836 0.214402086 0.3124318   0.1748077   0.3843143   0.8765256
#s[2]      0.3691174 0.212062373 0.3035039   0.1519308   0.1930381   0.2066811
#sigma     3.1524819 0.003510676 0.1739904   2.8400096   3.0331962   3.1411533
#                75%       97.5%       n_eff     Rhat
#h1      131.0000000 131.0000000 8000.000000      NaN
#h_theta 123.0556379 124.3928800  540.453594 1.002660
#theta     6.8790801   7.3376348  679.024115 1.002296
#s[1]      0.9516115   0.9955989    2.123501 3.866466
#s[2]      0.2954409   0.9852540    2.048336 6.072550
#sigma     3.2635849   3.5204101 2456.231113 1.001078

So what does this mean? From the Rhat values for s[1] and s[2] you can see that there are issues with convergence for these two parameters. This is due to the fact that s[1] and s[2] are indistinguishable: they cannot be estimated both at the same time. A stronger regularising prior on s[1] and s[2] would probably drive one of the s parameters to zero.

I'm not sure I understand the point of s[1] and s[2]. From a statistical modelling point of view, you cannot obtain estimates for both parameters in the simple non-linear growth model that we're considering.

Update

As promised here is an update. This is turning into quite a long post, I've tried to make things as clear as possible by adding additional explanations.

Preliminary comments

Using positive_ordered as data type for s makes a significant difference in terms of convergence of solutions. It's not clear to me why that is the case, nor do I know how Stan implements positive_ordered, but it works.
In this hierarchical approach, we partially pool height data across all boys by considering h1 ~ normal(mu_h1, sigma_h1), with priors on the hyperparameters mu_h1 ~ normal(max(y), 10) and a half-Cauchy prior on sigma_h1 ~ cauchy(0, 10) (it's half-Cauchy because sigma is declared as real<lower=0>).
To be honest, I am unsure about the interpretation (and interpretability) of some of the parameters. Estimates for h_1 and h_theta are very similar, and in some way cancel each other out. I would imagine that this creates some convergence issues when fitting the model, but as you can see further down, Rhat values seem ok. Still, as I don't know enough about the model, data and its context, I remain somewhat skeptical as to the interpretability of those parameters. Extending the model by turning some of the other parameters into group-level parameters is straightforward from a statistical modelling point of view; however I imagine that difficulties will arise from the indistinguishability and lack of interpretability.

All these points aside, this should give you a practical example of how to implement a hierarchical model.

The Stan model

model_code <- "
data {
    int N;                                       // Number of observations
    int J;                                       // Number of boys
    int<lower=1,upper=J> boy[N];                 // Boy of observation
    real y[N];                                   // Height
    real t[N];                                   // Time
}

parameters {
    real<lower=0> h1[J];
    real<lower=0> h_theta;
    real<lower=0> theta;
    positive_ordered[2] s;
    real<lower=0> sigma;

    // Hyperparameters
    real<lower=0> mu_h1;
    real<lower=0> sigma_h1;
}

transformed parameters {
    vector[N] mu;

    for (i in 1:N)
        mu[i] = h1[boy[i]] - 2 * (h1[boy[i]] - h_theta) / (exp(s[1] * (t[i] - theta)) + (exp(s[2] * (t[i] - theta))));
}

model {
    h1 ~ normal(mu_h1, sigma_h1);          // Partially pool h1 parameters across boys

    mu_h1 ~ normal(max(y), 10);            // Prior on h1 hyperparameter mu
    sigma_h1 ~ cauchy(0, 10);              // Half-Cauchy prior on h1 hyperparameter sigma
    h_theta ~ normal(max(y), 2);           // Prior on h_theta
    theta ~ normal(max(t), 2);             // Prior on theta
    s ~ cauchy(0, 1);                      // Half-Cauchy priors on s[1] and s[2]

    y ~ normal(mu, sigma);
}
"

To summarise: We pool height data from all boys to improve estimates at the group (i.e. boy) level, by modelling the adult height parameter as h1 ~ normal(mu_h1, sigma_h1), where the hyperparameters mu_h1 and sigma_h1 characterise the normal distribution of adult height values across all boys. We choose weakly informative priors on the hyperparameters, and choose weakly informative priors on all remaining parameters similar to the first complete-pooling example.

Fit the model

# Fit model
fit <- stan(
    model_code = model_code,
    data = list(
        N = nrow(df),
        J = length(unique(df$boy)),
        boy = df$boy,
        y = df$height,
        t = df$age),
    iter = 4000)

Extract summary

We extract parameter estimates for all parameters; note that we now have as many h1 parameters as there are groups (i.e. boys).

# Get summary
summary(fit, pars = c("h1", "h_theta", "theta", "s", "sigma"))$summary
#               mean      se_mean         sd        2.5%         25%        50%
#h1[1]   142.9406153 0.1046670943 2.41580757 138.4272280 141.2858391 142.909765
#h1[2]   143.7054020 0.1070466445 2.46570025 139.1301456 142.0233342 143.652657
#h1[3]   144.0352331 0.1086953809 2.50145442 139.3982034 142.3131167 143.971473
#h1[4]   143.8589955 0.1075753575 2.48015745 139.2689731 142.1666685 143.830347
#h1[5]   144.7359976 0.1109871908 2.55284812 140.0529359 142.9917503 144.660586
#h1[6]   143.9844938 0.1082691127 2.49497990 139.3378948 142.2919990 143.926931
#h1[7]   144.3857221 0.1092604239 2.51645359 139.7349112 142.6665955 144.314645
#h1[8]   143.7469630 0.1070594855 2.46860328 139.1748700 142.0660983 143.697302
#h1[9]   143.6841113 0.1072208284 2.47391295 139.0885987 141.9839040 143.644357
#h1[10]  142.9518072 0.1041206784 2.40729732 138.4289207 141.3114204 142.918407
#h1[11]  143.5352502 0.1064173663 2.45712021 138.9607665 141.8547610 143.483157
#h1[12]  143.0941582 0.1050061258 2.42894673 138.5579378 141.4295430 143.055576
#h1[13]  143.6194965 0.1068494690 2.46574352 138.9426195 141.9412820 143.577920
#h1[14]  143.4477182 0.1060254849 2.44776536 138.9142081 141.7708660 143.392231
#h1[15]  143.1415683 0.1049131998 2.42575487 138.6246642 141.5014391 143.102219
#h1[16]  143.5686919 0.1063594201 2.45328456 139.0064573 141.8962853 143.510276
#h1[17]  144.0170715 0.1080567189 2.49269747 139.4162885 142.3138300 143.965127
#h1[18]  143.4740997 0.1064867748 2.45545200 138.8768051 141.7989566 143.426211
#h_theta 134.3394366 0.0718785944 1.72084291 130.9919889 133.2348411 134.367152
#theta     8.2214374 0.0132434918 0.45236221   7.4609612   7.9127800   8.164685
#s[1]      0.1772044 0.0004923951 0.01165119   0.1547003   0.1705841   0.177522
#s[2]      1.6933846 0.0322953612 1.18334358   0.6516669   1.1630900   1.463148
#sigma     2.2601677 0.0034146522 0.13271459   2.0138514   2.1657260   2.256678
#                75%       97.5%     n_eff     Rhat
#h1[1]   144.4795105 147.8847202  532.7265 1.008214
#h1[2]   145.2395543 148.8419618  530.5599 1.008187
#h1[3]   145.6064981 149.2080965  529.6183 1.008087
#h1[4]   145.4202919 149.0105666  531.5363 1.008046
#h1[5]   146.3200407 150.0701757  529.0592 1.008189
#h1[6]   145.5551778 149.1365279  531.0372 1.008109
#h1[7]   145.9594956 149.5996605  530.4593 1.008271
#h1[8]   145.3032680 148.8695637  531.6824 1.008226
#h1[9]   145.2401743 148.7674840  532.3662 1.008023
#h1[10]  144.4811712 147.9218834  534.5465 1.007937
#h1[11]  145.1153635 148.5968945  533.1235 1.007988
#h1[12]  144.6479561 148.0546831  535.0652 1.008115
#h1[13]  145.1660639 148.6562172  532.5386 1.008138
#h1[14]  144.9975197 148.5273804  532.9900 1.008067
#h1[15]  144.6733010 148.1130207  534.6057 1.008128
#h1[16]  145.1163764 148.6027096  532.0396 1.008036
#h1[17]  145.5578107 149.2014363  532.1519 1.008052
#h1[18]  145.0249329 148.4886949  531.7060 1.008055
#h_theta 135.4870338 137.6753254  573.1698 1.006818
#theta     8.4812339   9.2516700 1166.7226 1.002306
#s[1]      0.1841457   0.1988365  559.9036 1.005333
#s[2]      1.8673249   4.1143099 1342.5839 1.001562
#sigma     2.3470429   2.5374239 1510.5824 1.001219

Visualise adult height estimates

Finally we plot adult height estimates h_1 for all boys, including 50% and 97% confidence intervals.

# Plot h1 values
summary(fit, pars = c("h1"))$summary %>%
    as.data.frame() %>%
    rownames_to_column("Variable") %>%
    mutate(
        Variable = gsub("(h1\\[|\\])", "", Variable),
        Variable = df$key[match(Variable, df$boy)]) %>%
    ggplot(aes(x = `50%`, y = Variable)) +
    geom_point(size = 3) +
    geom_segment(aes(x = `2.5%`, xend = `97.5%`, yend = Variable), size = 1) +
    geom_segment(aes(x = `25%`, xend = `75%`, yend = Variable), size = 2) +
    labs(x = "Median (plus/minus 95% and 50% CIs)", y = "h1")

来源：https://stackoverflow.com/questions/50749414/developing-hierarchical-version-of-nonlinear-growth-curve-model-in-stan

标签

statistics

stan

rstan