-
Notifications
You must be signed in to change notification settings - Fork 1
/
hierarchical_modelling.qmd
422 lines (346 loc) · 9.88 KB
/
hierarchical_modelling.qmd
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
---
title: "Hierarchical modelling"
subtitle: "SHARP Bayesian Modeling for Environmental Health Workshop"
author: "Theo Rashid, Elizaveta Semenova"
date: "August 14 2023"
format: html
---
```{r, echo=FALSE, warning=FALSE, message=FALSE}
library(here)
library(tidyverse)
library(nimble)
library(bayesplot)
library(posterior)
library(hrbrthemes)
library(sf)
library(colorspace)
extrafont::loadfonts()
theme_set(theme_ipsum())
color_scheme_set(scheme = "viridis")
set.seed(2)
```
## Goal of this computing lab session
This goal of this lab is to introduce hierarchical modelling using the `NIMBLE` modelling framework.
## What's going to happen in this lab session?
During this lab session, we will:
1. Write a hierarchical model in `NIMBLE`;
2. Compare the model to more basic non-hierarchical models; and
3. Discuss the advantages of hierarchical models.
## Introduction
In this task, we will explore mortality data in Italy during September 2018 using hierarchical modelling.
::: aside
The standardised mortality ratio (SMR) is the ratio of the number of deaths observed in a population over a given period to the number that would be expected over the same period if the study population had the same age-specific rates as the standard population.
:::
We will estimate the standardised mortality ratio (SMR) for populations in different provinces of Italy using three types of models:
1. Full pooling
2. No pooling
3. Partial pooling
::: aside
This workflow (fitting full/no/partial pooling models) is adapted from a classic Bayesian modelling example: the radon model. You can read about this model in Gelman and Hill's _Data Analysis Using Regression and Multilevel/Hierarchical Models_ (2006) or on the [tensorflow probability website](https://www.tensorflow.org/probability/examples/Multilevel_Modeling_Primer).
:::
## Exploratory data analysis
Let's load in the data.
```{r}
data <- read_rds(here("data", "italy", "italy_mortality.rds"))
glimpse(data)
summary(data)
```
Let's collapse the time dimension for now and focus on estimating the death rate in the final year and the 9th month.
```{r}
data <- data |>
# make dummies for each province, useful for modelling
mutate(provincia_id = data |> group_by(SIGLA) |> group_indices()) |>
filter(year == 2018) |>
filter(month == 9) |>
arrange(SIGLA)
```
Let's look at the distribution the number of deaths in each province.
```{r}
#| label: fig-deaths-province
#| warning: false
#| fig-height: 10
#| fig-cap: Number of deaths at each time point in each province.
data |>
ggplot(aes(x = SIGLA, y = deaths)) +
geom_jitter(size = 0.4, alpha = 0.8, colour = "darkcyan") +
coord_flip() +
theme(axis.text.y = element_text(size = 7))
```
Let's plot the mean number of deaths in that month in that province.
```{r}
shp_italy <- read_rds(here("data", "italy", "italy_shp.rds")) |> arrange(SIGLA)
shp_italy |>
left_join(
data |>
group_by(SIGLA) |>
summarise(mean_deaths = mean(deaths))
) |>
ggplot(aes(fill = mean_deaths)) +
geom_sf(colour = "white") +
scale_fill_continuous_sequential(palette = "Reds") +
theme_void()
```
We have a variable `expected`, which gives a good (if a slight underestimation) approximation to the number of deaths.
```{r}
data |>
ggplot(aes(x = expected, y = deaths)) +
geom_abline(slope = 1, intercept = 0, linewidth = 0.1) +
geom_point(size = 0.3)
```
```{r}
data |>
ggplot(aes(x = deaths / expected)) +
geom_histogram()
```
We are going to estimate the region-specific SMR.
This is the ratio of actual deaths to expected deaths in each region.
## Full pooling model
First, let's treat all regions the same and estimate a single national prevalence.
The model is as follows:
Priors
$$
\alpha \sim N(0, 5)
$$
Likelihood
$$
\begin{split}
y_i &\sim \text{Pois}(\mu_i) \quad i = 1,..., N \\
\log(\mu_i) &= \log(E_i) + \alpha
\end{split}
$$
The parameter $\alpha$ here is common for all observations $y_i$, regardless of what province the observation belongs to.
```{r}
constants <- list(
N = nrow(data),
Np = max(data$provincia_id),
province = data$provincia_id
)
inits <- list(list(alpha = 0), list(alpha = 0.1))
nimble_data <- list(
y = as.integer(data$deaths),
E = data$expected
)
```
```{r}
full_pooling_model <- nimbleCode({
# priors
alpha ~ dnorm(0, 5)
# likelihood
for (i in 1:N) {
y[i] ~ dpois(mu[i])
log(mu[i]) <- log(E[i]) + alpha
}
})
```
```{r}
full_pooling_samples <- nimbleMCMC(
code = full_pooling_model,
data = nimble_data,
constants = constants,
inits = inits,
monitors = c("alpha"),
niter = 2000,
nburnin = 1000,
thin = 1,
nchains = 2,
setSeed = 1,
samplesAsCodaMCMC = TRUE,
progressBar = TRUE
)
```
Follow the Bayesian workflow: look at MCMC outputs, check convergence using r-hat.
```{r}
summarise_draws(full_pooling_samples, default_summary_measures())
summarise_draws(full_pooling_samples, default_convergence_measures())
```
Let's plot the SMR in each province.
```{r}
full_pooling_posterior <- as_draws_array(full_pooling_samples)
SMR_full_pooling <- exp(full_pooling_posterior[, , "alpha"]) |>
apply(MARGIN = 3, FUN = median)
p_full <- shp_italy |>
mutate(SMR = SMR_full_pooling) |>
ggplot(aes(fill = SMR)) +
geom_sf(colour = "white") +
scale_fill_continuous_sequential(palette = "Reds") +
theme_void()
p_full
```
Pretty boring, eh?
## No pooling model
Here, we treat all regions separately and estimate the SMR in each region separately with fixed effects.
Priors
$$
\alpha_j \sim N(0, 1) \quad j = 1,..., N_p
$$
Likelihood
$$
\begin{split}
y_i &\sim \text{Pois}(\mu_i) \quad i = 1,..., N \\
\log(\mu_i) &= \log(E_i) + \alpha_{j[i]}
\end{split}
$$
There is now an $\alpha$ for each province, so the map will no longer be uniform.
```{r}
inits <- list(
list(
alpha = rep(0, constants$Np)
),
list(
alpha = rep(0.1, constants$Np)
)
)
```
```{r}
no_pooling_model <- nimbleCode({
# priors
for (j in 1:Np) {
alpha[j] ~ dnorm(0, sd = 1)
}
# likelihood
for (i in 1:N) {
y[i] ~ dpois(mu[i])
log(mu[i]) <- log(E[i]) + alpha[province[i]]
}
})
```
```{r}
no_pooling_samples <- nimbleMCMC(
code = no_pooling_model,
data = nimble_data,
constants = constants,
inits = inits,
monitors = c("alpha"),
niter = 2000,
nburnin = 1000,
thin = 1,
nchains = 2,
setSeed = 1,
samplesAsCodaMCMC = TRUE,
progressBar = TRUE
)
```
```{r}
summarise_draws(no_pooling_samples, default_summary_measures())
summarise_draws(no_pooling_samples, default_convergence_measures())
```
The SMR in each province
```{r}
no_pooling_posterior <- as_draws_array(no_pooling_samples)
SMR_no_pooling <- exp(no_pooling_posterior[, , 1:107]) |>
apply(MARGIN = 3, FUN = median)
p_no <- shp_italy |>
mutate(SMR = SMR_no_pooling) |>
ggplot(aes(fill = SMR)) +
geom_sf(colour = "white") +
scale_fill_continuous_sequential(palette = "Reds") +
theme_void()
p_no
```
## Partial pooling model
This is also known as a varying intercept model.
We fit a random effect for each province.
The size of the random effect is controlled by the hyperparameter $\sigma_p$.
There is now a hierarchy in the parameters between $\sigma_p$ and the effects for each province $\theta$ – hence, "hierarchical modelling".
Priors
$$
\begin{split}
\alpha &\sim N(0,5), \\
\sigma_p &\sim N^+(1) \\
\theta_j &\sim N(0, \sigma^2_p) \quad j = 1,..., N_p
\end{split}
$$
Likelihood
$$
\begin{split}
y_i &\sim \text{Pois}(\mu_i) \quad i = 1,..., N \\
\log(\mu_i) &= \log(E_i) + \alpha + \theta_{j[i]}
\end{split}
$$
```{r}
inits <- list(
list(
alpha = 0,
theta = rep(0, constants$Np),
sigma_p = 1
),
list(
alpha = 0.1,
theta = rep(0, constants$Np),
sigma_p = 1
)
)
```
```{r}
partial_pooling_model <- nimbleCode({
# priors
alpha ~ dnorm(0, 5)
sigma_p ~ T(dnorm(0, 1), 0, Inf) # half-normal
for (j in 1:Np) {
theta[j] ~ dnorm(0, sd = sigma_p)
}
# likelihood
for (i in 1:N) {
y[i] ~ dpois(mu[i])
log(mu[i]) <- log(E[i]) + alpha + theta[province[i]]
}
})
```
```{r}
partial_pooling_samples <- nimbleMCMC(
code = partial_pooling_model,
data = nimble_data,
constants = constants,
inits = inits,
monitors = c("alpha", "sigma_p", "theta"),
niter = 2000,
nburnin = 1000,
thin = 1,
nchains = 2,
setSeed = 1,
samplesAsCodaMCMC = TRUE,
progressBar = TRUE
)
```
```{r}
summarise_draws(partial_pooling_samples, default_summary_measures())
summarise_draws(partial_pooling_samples, default_convergence_measures())
```
With more complicated models, sometimes it's nice to see if the traceplots are working too.
```{r}
mcmc_trace(partial_pooling_samples, regex_pars = "alpha")
```
There is a little bit more manipulation required to estimate the SMR for the more complicated model.
```{r}
partial_pooling_posterior <- as_draws_array(partial_pooling_samples)
SMR_partial_pooling <- exp(
sweep(
partial_pooling_posterior[, , 3:109],
partial_pooling_posterior[, , 1],
MARGIN = c(1, 2),
FUN = "+"
)
) |>
apply(MARGIN = 3, FUN = median)
p_partial <- shp_italy |>
mutate(SMR = SMR_partial_pooling) |>
ggplot(aes(fill = SMR)) +
geom_sf(colour = "white") +
scale_fill_continuous_sequential(palette = "Reds") +
theme_void()
p_partial
```
## Comparing each of the models
Look at how each of the models alters the fit.
```{r}
p_full + labs(caption = "Full pooling model")
p_no + labs(caption = "No pooling model")
p_partial + labs(caption = "Partial pooling model")
```
There is clearly more smoothing in the partial pooling model.
## Closing remarks
In this lab session, we have explored how to fit a hierarchical model in `NIMBLE`.
We compared a model where the spatial effects are not pooled, fully pooled, and finally partially pooled.
Hierarchical models are extremely useful in practice.
We can use the structure of the problem to share information between similar populations.
For example, as well as neighbouring spatial units, we could borrow strength over similar age groups, adjacent time periods, or even diseases with similar aetiologies.